diff --git a/CMakeLists.txt b/CMakeLists.txt
index 19cb74ee0..534f6d85a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -469,6 +469,8 @@ set_target_properties(ggml_v3 PROPERTIES POSITION_INDEPENDENT_CODE ON)
 add_library(common2
             common/common.cpp
             common/common.h
+            common/json.cpp
+            common/json.h
             common/speculative.cpp
             common/speculative.h
             common/sampling.cpp
diff --git a/Makefile b/Makefile
index c44ee4557..0e1829752 100644
--- a/Makefile
+++ b/Makefile
@@ -110,10 +110,10 @@ endif
 CUBLASLD_FLAGS =
 CUBLAS_OBJS =
 
-OBJS_FULL += ggml-alloc.o ggml-cpu-traits.o ggml-quants.o ggml-cpu-quants.o kcpp-quantmapper.o kcpp-repackmapper.o unicode.o unicode-common.o unicode-data.o ggml-threading.o ggml-cpu-cpp.o gguf.o sgemm.o common.o speculative.o llama-impl.o sampling.o budget.o kcpputils.o kcppllmutils.o mtmdaudio.o ggml-rpc.o transport.o hash.o
-OBJS_SIMPLE += ggml-alloc.o ggml-cpu-traits.o ggml-quants_noavx2.o ggml-cpu-quants.o kcpp-quantmapper_noavx2.o kcpp-repackmapper_noavx2.o unicode.o unicode-common.o unicode-data.o ggml-threading.o ggml-cpu-cpp.o gguf.o sgemm_noavx2.o common.o speculative.o llama-impl.o sampling.o budget.o kcpputils.o kcppllmutils.o mtmdaudio.o ggml-rpc.o transport.o hash.o
-OBJS_SIMPLER += ggml-alloc.o ggml-cpu-traits.o ggml-quants_noavx1.o ggml-cpu-quants.o kcpp-quantmapper_noavx1.o kcpp-repackmapper_noavx1.o unicode.o unicode-common.o unicode-data.o ggml-threading.o ggml-cpu-cpp.o gguf.o sgemm_noavx1.o common.o speculative.o llama-impl.o sampling.o budget.o kcpputils.o kcppllmutils.o mtmdaudio.o ggml-rpc.o transport.o hash.o
-OBJS_FAILSAFE += ggml-alloc.o ggml-cpu-traits.o ggml-quants_failsafe.o ggml-cpu-quants.o kcpp-quantmapper_failsafe.o kcpp-repackmapper_failsafe.o unicode.o unicode-common.o unicode-data.o ggml-threading.o ggml-cpu-cpp.o gguf.o sgemm_failsafe.o common.o speculative.o llama-impl.o sampling.o budget.o kcpputils.o kcppllmutils.o mtmdaudio.o ggml-rpc.o transport.o hash.o
+OBJS_FULL += ggml-alloc.o ggml-cpu-traits.o ggml-quants.o ggml-cpu-quants.o kcpp-quantmapper.o kcpp-repackmapper.o unicode.o unicode-common.o unicode-data.o ggml-threading.o ggml-cpu-cpp.o gguf.o sgemm.o common.o common-json.o speculative.o llama-impl.o sampling.o budget.o kcpputils.o kcppllmutils.o mtmdaudio.o ggml-rpc.o transport.o hash.o
+OBJS_SIMPLE += ggml-alloc.o ggml-cpu-traits.o ggml-quants_noavx2.o ggml-cpu-quants.o kcpp-quantmapper_noavx2.o kcpp-repackmapper_noavx2.o unicode.o unicode-common.o unicode-data.o ggml-threading.o ggml-cpu-cpp.o gguf.o sgemm_noavx2.o common.o common-json.o speculative.o llama-impl.o sampling.o budget.o kcpputils.o kcppllmutils.o mtmdaudio.o ggml-rpc.o transport.o hash.o
+OBJS_SIMPLER += ggml-alloc.o ggml-cpu-traits.o ggml-quants_noavx1.o ggml-cpu-quants.o kcpp-quantmapper_noavx1.o kcpp-repackmapper_noavx1.o unicode.o unicode-common.o unicode-data.o ggml-threading.o ggml-cpu-cpp.o gguf.o sgemm_noavx1.o common.o common-json.o speculative.o llama-impl.o sampling.o budget.o kcpputils.o kcppllmutils.o mtmdaudio.o ggml-rpc.o transport.o hash.o
+OBJS_FAILSAFE += ggml-alloc.o ggml-cpu-traits.o ggml-quants_failsafe.o ggml-cpu-quants.o kcpp-quantmapper_failsafe.o kcpp-repackmapper_failsafe.o unicode.o unicode-common.o unicode-data.o ggml-threading.o ggml-cpu-cpp.o gguf.o sgemm_failsafe.o common.o common-json.o speculative.o llama-impl.o sampling.o budget.o kcpputils.o kcppllmutils.o mtmdaudio.o ggml-rpc.o transport.o hash.o
 
 # OS specific
 ifeq ($(UNAME_S),Linux)
@@ -673,6 +673,8 @@ llama-model.o: src/llama-model.cpp src/llama-model.h src/models/models.h ggml/in
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 common.o: common/common.cpp common/common.h common/log.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
+common-json.o: common/json.cpp common/json.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 speculative.o: common/speculative.cpp common/speculative.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 sampling.o: common/sampling.cpp common/common.h common/sampling.h common/log.h
diff --git a/common/arg.cpp b/common/arg.cpp
index 83c097116..f10aea7a5 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -5,6 +5,7 @@
 #include "common.h"
 #include "download.h"
 #include "json-schema-to-grammar.h"
+#include "json.h"
 #include "llama.h"
 #include "log.h"
 #include "sampling.h"
@@ -22,9 +23,6 @@
 #include <shellapi.h>
 #endif
 
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
-
 #include <algorithm>
 #include <cinttypes>
 #include <climits>
@@ -33,6 +31,7 @@
 #include <filesystem>
 #include <fstream>
 #include <list>
+#include <numeric>
 #include <regex>
 #include <set>
 #include <string>
@@ -56,7 +55,7 @@
 
 #define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 using namespace common_arg_utils;
 
 static std::initializer_list<enum llama_example> mmproj_examples = {
diff --git a/common/chat-auto-parser-generator.cpp b/common/chat-auto-parser-generator.cpp
index af84ff323..d7e117e4d 100644
--- a/common/chat-auto-parser-generator.cpp
+++ b/common/chat-auto-parser-generator.cpp
@@ -5,13 +5,12 @@
 #include "common.h"
 #include "json-schema-to-grammar.h"
 #include "log.h"
-#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <stdexcept>
 #include <string>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 // Helper to iterate over tools/functions
 static void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {
@@ -391,7 +390,7 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
 
         std::set<std::string> required;
         if (params.contains("required")) {
-            params.at("required").get_to(required);
+            required = params.at("required").get<std::set<std::string>>();
         }
 
         auto schema_info = common_schema_info();
diff --git a/common/chat-auto-parser-helpers.cpp b/common/chat-auto-parser-helpers.cpp
index 81b17e5e1..b37906bdf 100644
--- a/common/chat-auto-parser-helpers.cpp
+++ b/common/chat-auto-parser-helpers.cpp
@@ -4,14 +4,11 @@
 #include "chat-peg-parser.h"
 #include "chat.h"
 #include "log.h"
-#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <cctype>
 #include <numeric>
 
-using json = nlohmann::ordered_json;
-
 std::string trim_whitespace(const std::string & str) {
     size_t start = 0;
     while (start < str.length() && std::isspace(static_cast<unsigned char>(str[start]))) {
diff --git a/common/chat-auto-parser.h b/common/chat-auto-parser.h
index 074216b11..8ae15c91e 100644
--- a/common/chat-auto-parser.h
+++ b/common/chat-auto-parser.h
@@ -4,7 +4,7 @@
 #include "common.h"
 #include "jinja/caps.h"
 #include "peg-parser.h"
-#include "nlohmann/json.hpp"
+#include "json.h"
 
 #include <chrono>
 #include <optional>
@@ -12,7 +12,7 @@
 #include <utility>
 #include <vector>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 class common_chat_peg_builder;
 
diff --git a/common/chat-diff-analyzer.cpp b/common/chat-diff-analyzer.cpp
index d6d2af2d5..a7e370578 100644
--- a/common/chat-diff-analyzer.cpp
+++ b/common/chat-diff-analyzer.cpp
@@ -4,11 +4,11 @@
 #include "chat.h"
 #include "common.h"
 #include "log.h"
-#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <algorithm>
 #include <cctype>
+#include <numeric>
 #include <ostream>
 #include <sstream>
 
@@ -17,7 +17,7 @@
 #define ANSI_ORANGE "\033[1m\x1b[38;5;214m"
 #define ANSI_RED    "\033[1m\x1b[38;5;196m"
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 namespace autoparser {
 
@@ -929,7 +929,7 @@ void analyze_tools::analyze_tool_call_format_json_native(const std::string & cle
     int  json_end       = clean_haystack.find_last_of('}');
     std::string cut     = clean_haystack.substr(json_start, json_end - json_start + 1);
     json call_struct    = json::parse(cut);
-    auto register_field = [&](const std::string & prefix, const nlohmann::detail::iteration_proxy_value<json::iterator> & subel) {
+    auto register_field = [&](const std::string & prefix, const common_json_entry & subel) {
         if (subel.value().is_string() && std::string(subel.value()).find("call0000") != std::string::npos) {
             format.id_field = !prefix.empty() ? prefix + "." + subel.key() : subel.key();
         } else if (subel.value().is_string() && std::string(subel.value()) == fun_name_needle) {
diff --git a/common/chat-peg-parser.cpp b/common/chat-peg-parser.cpp
index 06737b165..79b97a80f 100644
--- a/common/chat-peg-parser.cpp
+++ b/common/chat-peg-parser.cpp
@@ -4,12 +4,10 @@
 #include "ggml.h"
 #include "peg-parser.h"
 
-#include <nlohmann/json.hpp>
-
 #include <cstdint>
 #include <functional>
 
-using ordered_json = nlohmann::ordered_json;
+using ordered_json = common_json;
 
 static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
     int count = 0;
diff --git a/common/chat-peg-parser.h b/common/chat-peg-parser.h
index 5d764dbaa..114fa049f 100644
--- a/common/chat-peg-parser.h
+++ b/common/chat-peg-parser.h
@@ -128,7 +128,7 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     // parameters_order: order in which JSON fields should be parsed
     common_peg_parser standard_json_tools(const std::string &              section_start,
                                           const std::string &              section_end,
-                                          const nlohmann::ordered_json &   tools,
+                                          const common_json &   tools,
                                           bool                             parallel_tool_calls,
                                           bool                             force_tool_calls,
                                           const std::string &              name_key = "",
@@ -143,13 +143,13 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     // Legacy-compatible helper for building XML/tagged style tool calls
     // Used by tests and manual parsers
     common_peg_parser standard_constructed_tools(const std::map<std::string, std::string> & markers,
-                                                 const nlohmann::ordered_json &             tools,
+                                                 const common_json &             tools,
                                                  bool                                       parallel_tool_calls,
                                                  bool                                       force_tool_calls);
 
     // Helper for Python-style function call format: name(arg1="value1", arg2=123)
     // Used by LFM2 and similar templates
-    common_peg_parser python_style_tool_calls(const nlohmann::ordered_json & tools,
+    common_peg_parser python_style_tool_calls(const common_json & tools,
                                               bool                           parallel_tool_calls,
                                               bool                           allow_json_literals);
 
@@ -158,19 +158,19 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     common_peg_parser python_or_json_value();
 
     // Implementation helpers for standard_json_tools — one per JSON tool call layout mode
-    common_peg_parser build_json_tools_function_is_key(const nlohmann::ordered_json & tools,
+    common_peg_parser build_json_tools_function_is_key(const common_json & tools,
                                                        const std::string &            args_key,
                                                        const std::string &            effective_args_key,
                                                        const std::string &            call_id_key,
                                                        const std::string &            gen_call_id_key);
 
-    common_peg_parser build_json_tools_nested_keys(const nlohmann::ordered_json & tools,
+    common_peg_parser build_json_tools_nested_keys(const common_json & tools,
                                                    const std::string &            effective_name_key,
                                                    const std::string &            effective_args_key,
                                                    const std::string &            call_id_key,
                                                    const std::string &            gen_call_id_key);
 
-    common_peg_parser build_json_tools_flat_keys(const nlohmann::ordered_json &   tools,
+    common_peg_parser build_json_tools_flat_keys(const common_json &   tools,
                                                  const std::string &              effective_name_key,
                                                  const std::string &              effective_args_key,
                                                  const std::string &              call_id_key,
diff --git a/common/chat.cpp b/common/chat.cpp
index 2fa9d2e07..bc133b55e 100644
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -6,6 +6,7 @@
 #include "common.h"
 #include "ggml.h"
 #include "json-schema-to-grammar.h"
+#include "json.h"
 #include "log.h"
 #include "reasoning-budget.h"
 #include "chat-auto-parser-generator.cpp"
@@ -30,6 +31,7 @@
 #include <ctime>
 #include <exception>
 #include <functional>
+#include <iomanip>
 #include <map>
 
 #include <optional>
@@ -39,7 +41,7 @@
 #include <utility>
 #include <vector>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
     auto               time       = std::chrono::system_clock::to_time_t(now);
@@ -57,7 +59,7 @@ static json safe_args_parse(const std::string & to_parse) {
     }
     try {
         return json::parse(stripped);
-    } catch (json::exception & e) {
+    } catch (const common_json_error & e) {
         return stripped;
     }
 }
@@ -497,17 +499,17 @@ struct messages_inp_normalizer {
         json normalized = json::array();
         for (const auto & msg : messages) {
             json copy = msg;
-            auto it = copy.find("content");
-            if (it != copy.end()) {
-                if (only_typed && it->is_string()) {
-                    *it = json::array({
+            if (copy.contains("content")) {
+                json & it = copy.at("content");
+                if (only_typed && it.is_string()) {
+                    it = json::array({
                         json{
                             {"type", "text"},
-                            {"text", it->get<std::string>()},
+                            {"text", it.get<std::string>()},
                         }
                     });
-                } else if (only_string && it->is_array()) {
-                    *it = concat_content_parts(*it);
+                } else if (only_string && it.is_array()) {
+                    it = concat_content_parts(it);
                 }
             }
             normalized.push_back(std::move(copy));
@@ -622,7 +624,7 @@ std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & too
 #include "peg-parser.cpp"
 #include "chat-peg-parser.cpp"
 
-common_chat_continuation common_chat_continuation_parse(const nlohmann::ordered_json & value) {
+common_chat_continuation common_chat_continuation_parse(const common_json & value) {
     if (value.is_boolean() && value.get<bool>()) {
         return COMMON_CHAT_CONTINUATION_AUTO;
     }
@@ -934,7 +936,7 @@ static void foreach_parameter(const json &
     const auto &          props = params.at("properties");
     std::set<std::string> required;
     if (params.contains("required") && params.at("required").is_array()) {
-        params.at("required").get_to(required);
+        required = params.at("required").get<std::set<std::string>>();
     }
     for (const auto & [name, prop] : props.items()) {
         bool is_required = (required.find(name) != required.end());
@@ -951,7 +953,7 @@ static std::string common_chat_template_direct_apply_impl(
     jinja::context ctx(tmpl.source());
 
     // messages_override is already built for this template, do not touch its content parts
-    nlohmann::ordered_json inp = nlohmann::ordered_json{
+    json inp = json{
         {"messages", messages_override.has_value()
             ? *messages_override
             : messages_inp_normalizer(tmpl.original_caps()).normalize(inputs.messages)},
@@ -1072,7 +1074,7 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
                 });
             } else if (msg.at("content").is_array()) {
                 auto blocks = msg.at("content");
-                content.insert(content.end(), blocks.begin(), blocks.end());
+                content.insert(blocks);
             }
         }
 
@@ -2252,7 +2254,7 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
 
                 std::set<std::string> required;
                 if (params.contains("required")) {
-                    params.at("required").get_to(required);
+                    required = params.at("required").get<std::set<std::string>>();
                 }
 
                 auto schema_info = common_schema_info();
@@ -2874,7 +2876,7 @@ static common_chat_params common_chat_params_init_minimax_m3(const common_chat_t
 
                 std::set<std::string> required;
                 if (schema.contains("required")) {
-                    schema.at("required").get_to(required);
+                    required = schema.at("required").get<std::set<std::string>>();
                 }
 
                 std::vector<common_peg_parser> required_elements;
@@ -2986,10 +2988,10 @@ static void system_message_not_supported(json & messages) {
             auto & second_msg = messages[1];
             second_msg["content"] = first_msg.at("content").get<std::string>()
                 + "\n" + second_msg.at("content").get<std::string>();
-            messages.erase(messages.begin());
+            messages.erase(0);
         } else {
             LOG_WRN("Removing system prompt due to template not supporting system role\n");
-            messages.erase(messages.begin());
+            messages.erase(0);
         }
     }
 }
diff --git a/common/chat.h b/common/chat.h
index 6d5b220ae..cb39e3458 100644
--- a/common/chat.h
+++ b/common/chat.h
@@ -8,7 +8,7 @@
 #include "jinja/runtime.h"
 #include "jinja/caps.h"
 
-#include "nlohmann/json_fwd.hpp"
+#include "json.h"
 
 #include <chrono>
 #include <functional>
@@ -17,7 +17,6 @@
 #include <vector>
 
 using chat_template_caps = jinja::caps;
-using json = nlohmann::ordered_json;
 
 struct common_chat_templates;
 
@@ -87,7 +86,7 @@ struct common_chat_msg {
     std::string                               tool_name;
     std::string                               tool_call_id;
 
-    nlohmann::ordered_json to_json_oaicompat(bool concat_typed_text = false) const;
+    common_json to_json_oaicompat(bool concat_typed_text = false) const;
 
     std::string render_content(const std::string & delimiter = "\n\n") const;
 
@@ -211,7 +210,7 @@ struct common_chat_msg_delimiters {
     // split tokens into message spans. skips maps a start index to a length of a region to jump over without matching
     common_chat_msg_spans split(const llama_tokens & tokens, const std::map<size_t, size_t> & skips = {}) const;
 
-    nlohmann::ordered_json to_json() const;
+    common_json to_json() const;
 };
 
 struct common_chat_tool {
@@ -350,16 +349,16 @@ common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::strin
 bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates);
 
 // Parses a JSON array of messages in OpenAI's chat completion API format.
-std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
+std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const common_json & messages);
 
-std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
+std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const common_json & tools);
 
-common_chat_continuation common_chat_continuation_parse(const nlohmann::ordered_json & value);
+common_chat_continuation common_chat_continuation_parse(const common_json & value);
 
 // DEPRECATED: only used in tests
-nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
+common_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
 
-nlohmann::ordered_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
+common_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
 
 // get template caps, useful for reporting to server /props endpoint
 std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates);
@@ -386,4 +385,4 @@ struct common_chat_prompt_preset {
 
 common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates * chat_templates);
 
-common_chat_msg_delimiters common_chat_msg_delimiters_parse(const nlohmann::ordered_json & delimiters);
+common_chat_msg_delimiters common_chat_msg_delimiters_parse(const common_json & delimiters);
diff --git a/common/common.cpp b/common/common.cpp
index c71068016..5f6acfa7e 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -408,10 +408,11 @@ void common_params_print_info(const common_params & params, bool print_devices)
 #endif
     COM_TRC("%s: build %d (%s) with %s for %s%s\n", __func__, llama_build_number(), llama_commit(), llama_compiler(), llama_build_target(), build_type);
 
-    COM_INF("%s: verbosity = %d (adjust with the `-lv N` CLI arg)\n", __func__, common_log_get_verbosity_thold());
+    const int verbosity = common_log_get_verbosity_thold();
+    COM_INF("%s: verbosity = %d (adjust with the `-lv N` CLI arg)\n", __func__, verbosity);
 
     // device enumeration creates a primary context on CUDA backends, skip it when the caller does not own any device
-    if (print_devices) {
+    if (print_devices && verbosity >= LOG_LEVEL_TRACE) {
         COM_TRC("%s", "device_info:\n");
         for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
             auto * dev = ggml_backend_dev_get(i);
diff --git a/common/download.cpp b/common/download.cpp
index 2509f75ab..4b28a708c 100644
--- a/common/download.cpp
+++ b/common/download.cpp
@@ -5,9 +5,7 @@
 #include "log.h"
 #include "download.h"
 #include "hf-cache.h"
-
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <algorithm>
 #include <filesystem>
@@ -44,8 +42,6 @@
 #include <unistd.h>
 #endif
 
-using json = nlohmann::ordered_json;
-
 //
 // downloader
 //
@@ -856,8 +852,8 @@ static std::string common_docker_get_token(const std::string & repo) {
         throw std::runtime_error("Failed to get Docker registry token, HTTP code: " + std::to_string(res.first));
     }
 
-    std::string            response_str(res.second.begin(), res.second.end());
-    nlohmann::ordered_json response = nlohmann::ordered_json::parse(response_str);
+    std::string response_str(res.second.begin(), res.second.end());
+    common_json response = common_json::parse(response_str);
 
     if (!response.contains("token")) {
         throw std::runtime_error("Docker registry token response missing 'token' field");
@@ -919,9 +915,9 @@ std::string common_docker_resolve_model(const std::string & docker) {
             throw std::runtime_error("Failed to get Docker manifest, HTTP code: " + std::to_string(manifest_res.first));
         }
 
-        std::string            manifest_str(manifest_res.second.begin(), manifest_res.second.end());
-        nlohmann::ordered_json manifest = nlohmann::ordered_json::parse(manifest_str);
-        std::string            gguf_digest;  // Find the GGUF layer
+        std::string manifest_str(manifest_res.second.begin(), manifest_res.second.end());
+        common_json manifest = common_json::parse(manifest_str);
+        std::string gguf_digest;  // Find the GGUF layer
         if (manifest.contains("layers")) {
             for (const auto & layer : manifest["layers"]) {
                 if (layer.contains("mediaType")) {
diff --git a/common/hf-cache.cpp b/common/hf-cache.cpp
index f1dacaa47..50d6dd610 100644
--- a/common/hf-cache.cpp
+++ b/common/hf-cache.cpp
@@ -4,9 +4,7 @@
 #include "common.h"
 #include "log.h"
 #include "http.h"
-
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <filesystem>
 #include <fstream>
@@ -15,8 +13,6 @@
 #include <string_view>
 #include <stdexcept>
 
-namespace nl = nlohmann;
-
 #if defined(_WIN32)
 #define WIN32_LEAN_AND_MEAN
 #ifndef NOMINMAX
@@ -195,8 +191,8 @@ static void safe_write_file(const fs::path & path, const std::string & data) {
     }
 }
 
-static nl::json api_get(const std::string & url,
-                        const std::string & token) {
+static common_json api_get(const std::string & url,
+                           const std::string & token) {
     auto [cli, parts] = common_http_client(url);
 
     httplib::Headers headers = {
@@ -214,10 +210,10 @@ static nl::json api_get(const std::string & url,
         auto body = res->body;
 
         if (res->status == 200) {
-            return nl::json::parse(res->body);
+            return common_json::parse(res->body);
         }
         try {
-            body = nl::json::parse(res->body)["error"].get<std::string>();
+            body = common_json::parse(res->body)["error"].get<std::string>();
         } catch (...) { }
 
         throw std::runtime_error("GET failed (" + std::to_string(res->status) + "): " + body);
@@ -280,7 +276,7 @@ static std::string get_repo_commit(const std::string & repo_id,
         safe_write_file(refs_path / name, commit);
         return commit;
 
-    } catch (const nl::json::exception & e) {
+    } catch (const common_json_error & e) {
         LOG_ERR("%s: JSON error: %s\n", __func__, e.what());
     } catch (const std::exception & e) {
         LOG_ERR("%s: error: %s\n", __func__, e.what());
@@ -358,7 +354,7 @@ hf_files get_repo_files(const std::string & repo_id,
 
             files.push_back(file);
         }
-    } catch (const nl::json::exception & e) {
+    } catch (const common_json_error & e) {
         LOG_ERR("%s: JSON error: %s\n", __func__, e.what());
     } catch (const std::exception & e) {
         LOG_ERR("%s: error: %s\n", __func__, e.what());
diff --git a/common/jinja/README.md b/common/jinja/README.md
index 829124076..5b97fc92c 100644
--- a/common/jinja/README.md
+++ b/common/jinja/README.md
@@ -7,7 +7,7 @@ The implementation can be found in the `common/jinja` directory.
 ## Key Features
 
 - Input marking: security against special token injection
-- Decoupled from `nlohmann::json`: this dependency is only used for JSON-to-internal type translation and is completely optional
+- Decoupled from the JSON library: `common_json` is only used for JSON-to-internal type translation and is completely optional
 - Minimal primitive types: int, float, bool, string, array, object, none, undefined
 - Detailed logging: allow source tracing on error
 - Clean architecture: workarounds are applied to input data before entering the runtime (see `common/chat.cpp`)
diff --git a/common/jinja/caps.cpp b/common/jinja/caps.cpp
index 5e6b34f76..567de5ba7 100644
--- a/common/jinja/caps.cpp
+++ b/common/jinja/caps.cpp
@@ -4,7 +4,7 @@
 
 // note: the json dependency is only for defining input in a convenient way
 // we can remove it in the future when we figure out a better way to define inputs using jinja::value
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <functional>
 #include <sstream>
@@ -14,7 +14,7 @@
 #endif
 #define FILENAME "jinja-caps"
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 namespace jinja {
 
diff --git a/common/jinja/value.cpp b/common/jinja/value.cpp
index 94f2aea56..07fb0d0e9 100644
--- a/common/jinja/value.cpp
+++ b/common/jinja/value.cpp
@@ -3,7 +3,7 @@
 #include "value.h"
 
 // for converting from JSON to jinja values
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <sstream>
 #include <string>
@@ -1358,7 +1358,7 @@ const func_builtins & value_undefined_t::get_builtins() const {
 //////////////////////////////////
 
 
-static value from_json(const nlohmann::ordered_json & j, bool mark_input) {
+static value from_json(const common_json & j, bool mark_input) {
     if (j.is_null()) {
         return mk_val<value_none>();
     } else if (j.is_boolean()) {
@@ -1455,7 +1455,7 @@ bool value_compare(const value & a, const value & b, value_compare_op op) {
 }
 
 template<>
-void global_from_json(context & ctx, const nlohmann::ordered_json & json_obj, bool mark_input) {
+void global_from_json(context & ctx, const common_json & json_obj, bool mark_input) {
     // printf("global_from_json: %s\n" , json_obj.dump(2).c_str());
     if (json_obj.is_null() || !json_obj.is_object()) {
         throw std::runtime_error("global_from_json: input JSON value must be an object");
diff --git a/common/jinja/value.h b/common/jinja/value.h
index 5cf85e4f5..4926fb680 100644
--- a/common/jinja/value.h
+++ b/common/jinja/value.h
@@ -86,7 +86,7 @@ struct context; // forward declaration
 // marking input can be useful for tracking data provenance
 // and preventing template injection attacks
 //
-// Note: T_JSON can be nlohmann::ordered_json
+// Note: T_JSON can be common_json
 template<typename T_JSON>
 void global_from_json(context & ctx, const T_JSON & json_obj, bool mark_input);
 
diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp
index 955b4e014..0aee51b26 100644
--- a/common/json-schema-to-grammar.cpp
+++ b/common/json-schema-to-grammar.cpp
@@ -1,9 +1,8 @@
 #include "json-schema-to-grammar.h"
 #include "common.h"
 
-#include <nlohmann/json.hpp>
-
 #include <algorithm>
+#include <limits>
 #include <map>
 #include <regex>
 #include <sstream>
@@ -12,7 +11,7 @@
 #include <unordered_set>
 #include <vector>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
     auto has_max = max_items != std::numeric_limits<int>::max();
@@ -917,7 +916,11 @@ public:
             return _add_rule(rule_name, _resolve_ref(schema["$ref"]));
         }
         if (schema.contains("oneOf") || schema.contains("anyOf")) {
-            std::vector<json> alt_schemas = schema.contains("oneOf") ? schema["oneOf"].get<std::vector<json>>() : schema["anyOf"].get<std::vector<json>>();
+            const json & alts = schema.contains("oneOf") ? schema.at("oneOf") : schema.at("anyOf");
+            std::vector<json> alt_schemas;
+            for (const auto & alt : alts) {
+                alt_schemas.push_back(alt);
+            }
             return _add_rule(rule_name, _generate_union_rule(name, alt_schemas));
         }
         if (schema_type.is_array()) {
@@ -1111,7 +1114,7 @@ common_schema_info::~common_schema_info() = default;
 common_schema_info::common_schema_info(common_schema_info &&) noexcept = default;
 common_schema_info & common_schema_info::operator=(common_schema_info &&) noexcept = default;
 
-void common_schema_info::resolve_refs(nlohmann::ordered_json & schema) {
+void common_schema_info::resolve_refs(common_json & schema) {
     impl_->resolve_refs(schema, "");
 }
 
@@ -1119,7 +1122,7 @@ void common_schema_info::resolve_refs(nlohmann::ordered_json & schema) {
 // Some models emit raw string values rather than JSON-encoded strings for string parameters.
 // If any branch of the schema (via oneOf, anyOf, $ref, etc.) permits a string, this returns
 // true, allowing callers to handle the value as a raw string for simplicity.
-bool common_schema_info::resolves_to_string(const nlohmann::ordered_json & schema) {
+bool common_schema_info::resolves_to_string(const common_json & schema) {
     std::unordered_set<std::string> visited_refs;
 
     std::function<bool(const json &)> check = [&](const json & s) -> bool {
@@ -1227,7 +1230,7 @@ bool common_schema_info::resolves_to_string(const nlohmann::ordered_json & schem
     return check(schema);
 }
 
-std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
+std::string json_schema_to_grammar(const common_json & schema, bool force_gbnf) {
 #ifdef LLAMA_USE_LLGUIDANCE
     if (!force_gbnf) {
         return "%llguidance {}\nstart: %json " + schema.dump();
@@ -1248,10 +1251,10 @@ std::string build_grammar(const std::function<void(const common_grammar_builder
         /* .add_rule = */ [&](const std::string & name, const std::string & rule) {
             return converter._add_rule(name, rule);
         },
-        /* .add_schema = */ [&](const std::string & name, const nlohmann::ordered_json & schema) {
+        /* .add_schema = */ [&](const std::string & name, const common_json & schema) {
             return converter.visit(schema, name == "root" ? "" : name);
         },
-        /* .resolve_refs = */ [&](nlohmann::ordered_json & schema) {
+        /* .resolve_refs = */ [&](common_json & schema) {
             converter.resolve_refs(schema, "");
         }
     };
diff --git a/common/json-schema-to-grammar.h b/common/json-schema-to-grammar.h
index 240d64231..84ed71c76 100644
--- a/common/json-schema-to-grammar.h
+++ b/common/json-schema-to-grammar.h
@@ -1,12 +1,12 @@
 #pragma once
 
-#include <nlohmann/json_fwd.hpp>
+#include "json.h"
 
 #include <functional>
 #include <memory>
 #include <string>
 
-std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
+std::string json_schema_to_grammar(const common_json & schema,
                                    bool force_gbnf = false);
 
 class common_schema_converter;
@@ -24,14 +24,14 @@ class common_schema_info {
     common_schema_info(common_schema_info &&) noexcept;
     common_schema_info & operator=(common_schema_info &&) noexcept;
 
-    void resolve_refs(nlohmann::ordered_json & schema);
-    bool resolves_to_string(const nlohmann::ordered_json & schema);
+    void resolve_refs(common_json & schema);
+    bool resolves_to_string(const common_json & schema);
 };
 
 struct common_grammar_builder {
     std::function<std::string(const std::string &, const std::string &)> add_rule;
-    std::function<std::string(const std::string &, const nlohmann::ordered_json &)> add_schema;
-    std::function<void(nlohmann::ordered_json &)> resolve_refs;
+    std::function<std::string(const std::string &, const common_json &)> add_schema;
+    std::function<void(common_json &)> resolve_refs;
 };
 
 struct common_grammar_options {
diff --git a/common/json.cpp b/common/json.cpp
new file mode 100644
index 000000000..37713cef2
--- /dev/null
+++ b/common/json.cpp
@@ -0,0 +1,433 @@
+#include "json.h"
+
+#include "ggml.h"
+
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
+
+#include <iterator>
+#include <new>
+#include <set>
+#include <unordered_map>
+#include <vector>
+
+using nlohmann::ordered_json;
+
+// a common_json is the backing value, so any value of a tree can be used as a common_json
+static_assert(sizeof(ordered_json)  <= sizeof(common_json),  "common_json storage is too small");
+static_assert(alignof(ordered_json) <= alignof(common_json), "common_json alignment is too weak");
+
+// runs fn and gives every error of the backing library as a common_json_error
+template <typename F>
+static decltype(auto) guard(F && fn) {
+    try {
+        return fn();
+    } catch (const ordered_json::exception & e) {
+        throw common_json_error(e.what());
+    }
+}
+
+static ordered_json & as_json(common_json * self) {
+    return *reinterpret_cast<ordered_json *>(self);
+}
+
+static const ordered_json & as_json(const common_json * self) {
+    return *reinterpret_cast<const ordered_json *>(self);
+}
+
+static common_json & as_common(ordered_json & json) {
+    return *reinterpret_cast<common_json *>(&json);
+}
+
+static const common_json & as_common(const ordered_json & json) {
+    return *reinterpret_cast<const common_json *>(&json);
+}
+
+static ordered_json to_json(const common_json_value & val) {
+    switch (val.type) {
+        case common_json_value::VAL_NULL:   return nullptr;
+        case common_json_value::VAL_BOOL:   return val.val_bool;
+        case common_json_value::VAL_INT:    return val.val_int;
+        case common_json_value::VAL_UINT:   return val.val_uint;
+        case common_json_value::VAL_DOUBLE: return val.val_double;
+        case common_json_value::VAL_STRING: return val.val_string;
+        case common_json_value::VAL_JSON:
+            // one owner means no one else can see this tree, so it is safe to move it out
+            // note: this makes a value single use, same as the json_ref of the backing library
+            if (val.val_json.use_count() == 1) {
+                return std::move(as_json(val.val_json.get()));
+            }
+            return as_json(val.val_json.get());
+    }
+
+    return nullptr;
+}
+
+common_json_value::common_json_value(const char * val) {
+    if (val) {
+        type       = VAL_STRING;
+        val_string = val;
+    } else {
+        type = VAL_NULL;
+    }
+}
+
+common_json_value::common_json_value(const common_json & val) :
+    type(VAL_JSON), val_json(std::make_shared<common_json>(val)) {}
+
+common_json_value::common_json_value(common_json && val) :
+    type(VAL_JSON), val_json(std::make_shared<common_json>(std::move(val))) {}
+
+// the ctors and get<T>() below are explicit specializations, giving strong symbols
+// an explicit instantiation is a weak symbol, dropped by some LTO builds (clang-cl)
+template <typename T>
+static std::shared_ptr<common_json> set_json(const std::set<T> & vals) {
+    common_json out = common_json::array();
+
+    for (const auto & val : vals) {
+        out.push_back(val);
+    }
+
+    return std::make_shared<common_json>(std::move(out));
+}
+
+// a set value is usable only for the types below
+#define COMMON_JSON_SET(...) template <> common_json_value::common_json_value(const std::set<__VA_ARGS__> & vals) : type(VAL_JSON), val_json(set_json(vals)) {}
+
+COMMON_JSON_SET(int)
+COMMON_JSON_SET(std::string)
+
+#undef COMMON_JSON_SET
+
+template <typename T>
+static std::shared_ptr<common_json> map_json(const T & vals) {
+    common_json out = common_json::object();
+
+    for (const auto & val : vals) {
+        out.set({ val.first, val.second });
+    }
+
+    return std::make_shared<common_json>(std::move(out));
+}
+
+// a map value is usable only for the types below
+#define COMMON_JSON_MAP(...) template <> common_json_value::common_json_value(const std::map<std::string, __VA_ARGS__> & vals) : type(VAL_JSON), val_json(map_json(vals)) {}
+
+COMMON_JSON_MAP(bool)
+COMMON_JSON_MAP(std::string)
+
+#undef COMMON_JSON_MAP
+
+// an unordered map value is usable only for the types below
+#define COMMON_JSON_UMAP(...) template <> common_json_value::common_json_value(const std::unordered_map<std::string, __VA_ARGS__> & vals) : type(VAL_JSON), val_json(map_json(vals)) {}
+
+COMMON_JSON_UMAP(size_t)
+
+#undef COMMON_JSON_UMAP
+
+template <typename T>
+static std::shared_ptr<common_json> vec_json(const std::vector<T> & vals) {
+    common_json out = common_json::array();
+
+    for (const auto & val : vals) {
+        out.push_back(val);
+    }
+
+    return std::make_shared<common_json>(std::move(out));
+}
+
+// a vector value is usable only for the types below
+// note: std::vector<bool> is not here, its proxy reference does not convert
+#define COMMON_JSON_VEC(...) template <> common_json_value::common_json_value(const std::vector<__VA_ARGS__> & vals) : type(VAL_JSON), val_json(vec_json(vals)) {}
+
+COMMON_JSON_VEC(int)
+COMMON_JSON_VEC(unsigned char)
+COMMON_JSON_VEC(unsigned int)
+COMMON_JSON_VEC(long)
+COMMON_JSON_VEC(unsigned long)
+COMMON_JSON_VEC(long long)
+COMMON_JSON_VEC(unsigned long long)
+COMMON_JSON_VEC(float)
+COMMON_JSON_VEC(double)
+COMMON_JSON_VEC(std::string)
+COMMON_JSON_VEC(std::vector<float>)
+COMMON_JSON_VEC(common_json)
+
+#undef COMMON_JSON_VEC
+
+common_json_value::common_json_value(std::initializer_list<common_json_item> items) :
+    type(VAL_JSON), val_json(std::make_shared<common_json>(items)) {}
+
+// null, same as the backing library
+// operator[] turns it into an object, push_back() into an array
+common_json::common_json() {
+    new (storage) ordered_json();
+}
+
+common_json::common_json(const common_json & other) {
+    new (storage) ordered_json(as_json(&other));
+}
+
+common_json::common_json(common_json && other) noexcept {
+    new (storage) ordered_json(std::move(as_json(&other)));
+}
+
+common_json::common_json(std::initializer_list<common_json_item> items) {
+    new (storage) ordered_json(ordered_json::object());
+
+    for (const auto & item : items) {
+        set(item);
+    }
+}
+
+common_json::common_json(const common_json_value & val) {
+    new (storage) ordered_json(to_json(val));
+}
+
+common_json::common_json(std::nullptr_t) {
+    new (storage) ordered_json(nullptr);
+}
+
+common_json & common_json::operator=(common_json other) noexcept {
+    as_json(this).swap(as_json(&other));
+
+    return *this;
+}
+
+common_json::~common_json() {
+    as_json(this).~basic_json();
+}
+
+common_json common_json::parse(const std::string & text) {
+    try {
+        // the assignment moves the parsed tree in, it does not copy
+        common_json out;
+        as_json(&out) = ordered_json::parse(text);
+        return out;
+    } catch (const std::exception & e) {
+        throw common_json_error(e.what());
+    }
+}
+
+common_json common_json::parse_no_throw(const std::string & text) {
+    common_json out;
+    as_json(&out) = ordered_json::parse(text, nullptr, false);
+    return out;
+}
+
+bool common_json::is_discarded() const {
+    return as_json(this).is_discarded();
+}
+
+common_json common_json::array() {
+    common_json out;
+    as_json(&out) = ordered_json::array();
+    return out;
+}
+
+common_json common_json::array(std::initializer_list<common_json_value> vals) {
+    common_json out;
+    ordered_json & arr = as_json(&out);
+    arr = ordered_json::array();
+
+    for (const auto & val : vals) {
+        arr.push_back(to_json(val));
+    }
+
+    return out;
+}
+
+common_json common_json::object() {
+    common_json out;
+    as_json(&out) = ordered_json::object();
+    return out;
+}
+
+common_json common_json::object(std::initializer_list<common_json_item> items) {
+    return common_json(items);
+}
+
+common_json common_json::make(const common_json_value & val) {
+    return common_json(val);
+}
+
+bool common_json::is_null()           const { return as_json(this).is_null(); }
+bool common_json::is_object()         const { return as_json(this).is_object(); }
+bool common_json::is_array()          const { return as_json(this).is_array(); }
+bool common_json::is_string()         const { return as_json(this).is_string(); }
+bool common_json::is_boolean()        const { return as_json(this).is_boolean(); }
+bool common_json::is_number()         const { return as_json(this).is_number(); }
+bool common_json::is_number_integer() const { return as_json(this).is_number_integer(); }
+bool common_json::is_number_float()   const { return as_json(this).is_number_float(); }
+
+bool   common_json::empty() const { return as_json(this).empty(); }
+size_t common_json::size()  const { return as_json(this).size(); }
+
+bool common_json::contains(const std::string & key) const {
+    return as_json(this).contains(key);
+}
+
+bool common_json::operator==(const common_json_value & val) const {
+    // compare a tree in place, to_json() would copy it
+    if (val.type == common_json_value::VAL_JSON) {
+        return as_json(this) == as_json(val.val_json.get());
+    }
+    return as_json(this) == to_json(val);
+}
+
+bool common_json::operator!=(const common_json_value & val) const {
+    return !(*this == val);
+}
+
+common_json       & common_json::at(const std::string & key)       { return guard([&]() -> common_json       & { return as_common(as_json(this).at(key)); }); }
+const common_json & common_json::at(const std::string & key) const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(key)); }); }
+common_json       & common_json::at(size_t idx)                    { return guard([&]() -> common_json       & { return as_common(as_json(this).at(idx)); }); }
+const common_json & common_json::at(size_t idx)              const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(idx)); }); }
+
+common_json       & common_json::operator[](const std::string & key)       { return guard([&]() -> common_json       & { return as_common(as_json(this)[key]); }); }
+const common_json & common_json::operator[](const std::string & key) const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(key)); }); }
+common_json       & common_json::operator[](size_t idx)                    { return guard([&]() -> common_json       & { return as_common(as_json(this)[idx]); }); }
+const common_json & common_json::operator[](size_t idx)              const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(idx)); }); }
+
+common_json       & common_json::front()       { return as_common(as_json(this).front()); }
+const common_json & common_json::front() const { return as_common(as_json(this).front()); }
+common_json       & common_json::back()        { return as_common(as_json(this).back()); }
+const common_json & common_json::back()  const { return as_common(as_json(this).back()); }
+
+void common_json::clear() {
+    as_json(this).clear();
+}
+
+void common_json::erase(const std::string & key) {
+    guard([&] { as_json(this).erase(key); });
+}
+
+void common_json::erase(size_t idx) {
+    guard([&] { as_json(this).erase(idx); });
+}
+
+void common_json::assign(const common_json_value & val) {
+    as_json(this) = to_json(val);
+}
+
+void common_json::set(const common_json_item & item) {
+    guard([&] { as_json(this)[item.key] = to_json(item.val); });
+}
+
+void common_json::push_back(const common_json_value & val) {
+    guard([&] { as_json(this).push_back(to_json(val)); });
+}
+
+void common_json::push_back(std::initializer_list<common_json_item> items) {
+    common_json val(items);
+
+    guard([&] { as_json(this).push_back(std::move(as_json(&val))); });
+}
+
+size_t common_json::count(const std::string & key) const {
+    return as_json(this).count(key);
+}
+
+void common_json::insert(const common_json & vals) {
+    guard([&] {
+        ordered_json & self = as_json(this);
+
+        self.insert(self.end(), as_json(&vals).begin(), as_json(&vals).end());
+    });
+}
+
+std::string common_json::dump(int indent) const {
+    return guard([&] { return as_json(this).dump(indent); });
+}
+
+std::string common_json::dump_safe(int indent) const {
+    return as_json(this).dump(indent, ' ', false, ordered_json::error_handler_t::replace);
+}
+
+// an array is indexed directly, an object needs a walk from the start
+common_json & common_json::iterator::operator*() const {
+    return guard([&]() -> common_json & {
+        ordered_json & j = as_json(node);
+
+        if (j.is_object()) {
+            return as_common(std::next(j.begin(), idx).value());
+        }
+        if (j.is_array()) {
+            return as_common(j[idx]);
+        }
+
+        // a plain value gives itself once, same as the backing library
+        return *node;
+    });
+}
+
+std::string common_json::iterator::key() const {
+    return guard([&] { return std::next(as_json(node).begin(), idx).key(); });
+}
+
+common_json::iterator common_json::begin() const {
+    return iterator(const_cast<common_json *>(this), 0);
+}
+
+common_json::iterator common_json::end() const {
+    return iterator(const_cast<common_json *>(this), size());
+}
+
+// the keys follow the backing library: the index for an array, "" for a plain value
+common_json::items_view::entry common_json::items_view::iterator::operator*() const {
+    return guard([&]() -> entry {
+        ordered_json & j = as_json(node);
+
+        if (j.is_object()) {
+            auto it = std::next(j.begin(), idx);
+
+            return { it.key(), as_common(it.value()) };
+        }
+        if (j.is_array()) {
+            return { std::to_string(idx), as_common(j[idx]) };
+        }
+
+        return { std::string(), *node };
+    });
+}
+
+common_json::items_view common_json::items() const {
+    return items_view(const_cast<common_json *>(this), size());
+}
+
+// the backing library cannot build a common_json, so this one is just a copy
+template <> common_json common_json::get<common_json>() const {
+    return *this;
+}
+
+// get<T>() is usable only for the types below
+
+#define COMMON_JSON_GET(...) template <> __VA_ARGS__ common_json::get<__VA_ARGS__>() const { return guard([&] { return as_json(this).get<__VA_ARGS__>(); }); }
+
+COMMON_JSON_GET(bool)
+COMMON_JSON_GET(int)
+COMMON_JSON_GET(unsigned int)
+COMMON_JSON_GET(long)
+COMMON_JSON_GET(unsigned long)
+COMMON_JSON_GET(long long)
+COMMON_JSON_GET(unsigned long long)
+COMMON_JSON_GET(float)
+COMMON_JSON_GET(double)
+COMMON_JSON_GET(std::string)
+COMMON_JSON_GET(std::vector<float>)
+COMMON_JSON_GET(std::vector<std::string>)
+COMMON_JSON_GET(std::set<std::string>)
+COMMON_JSON_GET(std::vector<int>)
+COMMON_JSON_GET(std::vector<size_t>)
+COMMON_JSON_GET(std::unordered_map<std::string, size_t>)
+
+#undef COMMON_JSON_GET
+
+// must stay below the get<std::string> specialization
+common_json::operator std::string() const {
+    return get<std::string>();
+}
+
+std::string common_json::value(const std::string & key, const char * def) const {
+    return contains(key) ? at(key).get<std::string>() : std::string(def);
+}
diff --git a/common/json.h b/common/json.h
new file mode 100644
index 000000000..f3ad4edee
--- /dev/null
+++ b/common/json.h
@@ -0,0 +1,352 @@
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <initializer_list>
+#include <iterator>
+#include <map>
+#include <memory>
+#include <set>
+#include <stdexcept>
+#include <string>
+#include <string_view>
+#include <type_traits>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+// common_json, a thin wrapper around vendor json library
+// the underlay library is pimpl, we are using nlohmann::json for now
+//
+// many features of the library are deliberately left out, to keep this interface small and generic and to keep compile time down
+//
+// some main differences compared to nlohmann::json :
+// - object keys keep the order in which they are added
+// - errors are always throw as common_json_error
+// - obj.push_back({key, val}) is intentionally unsupported to avoid confusion with push_back on a vector; write it as obj[key] = val for clarity
+// - a braced pair in value position does not build, e.g. {"key", {"a", "b"}}; write array({"a", "b"}) where nlohmann made an array
+//
+// in doubt, search the code base for an existing usage example; do not add anything to this header unless absolutely necessary
+
+class common_json;
+
+// common_json_value holds a list of these, and each of them holds a value, so one must come first
+struct common_json_item;
+
+struct common_json_error : std::runtime_error {
+    using std::runtime_error::runtime_error;
+};
+
+// one value, tagged so that this header stays free of the backing library
+// note: a value that holds a tree is single use, the second use gives null
+struct common_json_value {
+    enum value_type {
+        VAL_NULL,
+        VAL_BOOL,
+        VAL_INT,
+        VAL_UINT,
+        VAL_DOUBLE,
+        VAL_STRING,
+        VAL_JSON,
+    };
+
+    value_type type = VAL_NULL;
+
+    union {
+        bool     val_bool;
+        int64_t  val_int;
+        uint64_t val_uint = 0;
+        double   val_double;
+    };
+
+    std::string                  val_string;
+    std::shared_ptr<common_json> val_json;
+
+    common_json_value(std::nullptr_t = nullptr) : type(VAL_NULL) {}
+    common_json_value(bool val) : type(VAL_BOOL), val_bool(val) {}
+    common_json_value(std::string val) : type(VAL_STRING), val_string(std::move(val)) {}
+    // without this a string_view lands on the common_json ctor below and recurses
+    common_json_value(std::string_view val) : type(VAL_STRING), val_string(val) {}
+    common_json_value(const char * val);
+    common_json_value(const common_json & val);
+    common_json_value(common_json && val);
+    // only for the types instantiated in json.cpp, the rest fails at link time
+    template <typename T> common_json_value(const std::vector<T> & vals);
+    // a set becomes an array, in the set's own order
+    template <typename T> common_json_value(const std::set<T> & vals);
+    // a map becomes an object, keyed in the map's own order
+    template <typename T> common_json_value(const std::map<std::string, T> & vals);
+    template <typename T> common_json_value(const std::unordered_map<std::string, T> & vals);
+
+    // nested object, e.g. {"fn", {{"name", "x"}}}
+    // note: a nested pair {"a", "b"} does not build, use common_json::array({"a", "b"}) for an array
+    common_json_value(std::initializer_list<common_json_item> items);
+
+    template <typename T, typename std::enable_if<std::is_integral<T>::value && !std::is_same<T, bool>::value, int>::type = 0>
+    common_json_value(T val) : type(std::is_signed<T>::value ? VAL_INT : VAL_UINT) {
+        if (std::is_signed<T>::value) {
+            val_int = (int64_t) val;
+        } else {
+            val_uint = (uint64_t) val;
+        }
+    }
+
+    template <typename T, typename std::enable_if<std::is_floating_point<T>::value, int>::type = 0>
+    common_json_value(T val) : type(VAL_DOUBLE), val_double((double) val) {}
+};
+
+struct common_json_item {
+    std::string       key;
+    common_json_value val;
+
+    template <typename T>
+    common_json_item(std::string key, T && val) :
+        key(std::move(key)), val(std::forward<T>(val)) {}
+
+    // a braced list cannot deduce T, so it needs its own overload
+    common_json_item(std::string key, std::initializer_list<common_json_item> items) :
+        key(std::move(key)), val(items) {}
+};
+
+// the types common_json_value holds on its own
+// anything else reaches its common_json ctor and recurses forever
+template <typename T> struct common_json_is_value : std::integral_constant<bool,
+    std::is_arithmetic<T>::value ||
+    std::is_same<T, std::nullptr_t>::value ||
+    std::is_same<T, std::string>::value ||
+    std::is_same<T, std::string_view>::value ||
+    std::is_same<T, char *>::value ||
+    std::is_same<T, const char *>::value ||
+    std::is_same<T, common_json>::value> {};
+
+template <typename T, typename A>
+struct common_json_is_value<std::vector<T, A>> : std::true_type {};
+
+template <typename T, typename C, typename A>
+struct common_json_is_value<std::set<T, C, A>> : std::true_type {};
+
+template <typename V, typename C, typename A>
+struct common_json_is_value<std::map<std::string, V, C, A>> : std::true_type {};
+
+template <typename V, typename H, typename E, typename A>
+struct common_json_is_value<std::unordered_map<std::string, V, H, E, A>> : std::true_type {};
+
+class common_json {
+  public:
+    common_json();
+    common_json(const common_json & other);
+    common_json(common_json && other) noexcept;
+    common_json(std::initializer_list<common_json_item> items);
+    common_json(const common_json_value & val);
+
+    // direct, a value would need two conversions in a row
+    common_json(std::nullptr_t);
+
+    // one step, so that "abc" or a vector can go straight into a common_json
+    template <typename T, typename std::enable_if<!std::is_same<typename std::decay<T>::type, common_json>::value &&
+                                                  !std::is_same<typename std::decay<T>::type, common_json_value>::value, int>::type = 0>
+    common_json(T && val) : common_json(common_json_value(std::forward<T>(val))) {
+        static_assert(common_json_is_value<typename std::decay<T>::type>::value,
+                      "no common_json_value ctor holds this type, add one instead of letting it recurse");
+    }
+
+    // by value, same as the backing library
+    // the right side is copied before the left side can invalidate it, e.g. msg["a"] = msg.at("b")
+    common_json & operator=(common_json other) noexcept;
+
+    ~common_json();
+
+    // throws common_json_error if the text is not valid JSON
+    static common_json parse(const std::string & text);
+
+    // gives a discarded value instead of throwing, check it with is_discarded()
+    static common_json parse_no_throw(const std::string & text);
+
+    bool is_discarded() const;
+
+    static common_json array();
+    static common_json array(std::initializer_list<common_json_value> vals);
+    static common_json object();
+    static common_json object(std::initializer_list<common_json_item> items);
+
+    // holds a single value, e.g. make("abc").dump() gives "\"abc\""
+    static common_json make(const common_json_value & val);
+
+    bool is_null()    const;
+    bool is_object()  const;
+    bool is_array()   const;
+    bool is_string()  const;
+    bool is_boolean() const;
+    bool is_number()  const;
+    bool is_number_integer() const;
+    bool is_number_float()   const;
+
+    bool   empty() const;
+    size_t size()  const;
+
+    bool contains(const std::string & key) const;
+
+    bool operator==(const common_json_value & val) const;
+    bool operator!=(const common_json_value & val) const;
+
+    // at() throws common_json_error if the key is missing, operator[] adds a null value instead
+    // note: a const operator[] cannot add, it throws like at()
+    common_json       & at(const std::string & key);
+    const common_json & at(const std::string & key) const;
+    common_json       & at(size_t idx);
+    const common_json & at(size_t idx) const;
+
+    common_json       & operator[](const std::string & key);
+    const common_json & operator[](const std::string & key) const;
+    common_json       & operator[](const char * key)       { return (*this)[std::string(key)]; }
+    const common_json & operator[](const char * key) const { return (*this)[std::string(key)]; }
+    common_json       & operator[](int idx)       { return (*this)[to_idx(idx)]; }
+    const common_json & operator[](int idx) const { return (*this)[to_idx(idx)]; }
+    common_json       & operator[](size_t idx);
+    const common_json & operator[](size_t idx) const;
+
+    common_json       & front();
+    const common_json & front() const;
+    common_json       & back();
+    const common_json & back()  const;
+
+    void clear();
+
+    void erase(const std::string & key);
+    void erase(size_t idx);
+
+    // only for the types instantiated in json.cpp, the rest fails at link time
+    template <typename T> T get() const;
+
+    // implicit get<T>() for plain values, so they can be assigned to their C++ type directly
+    // note: kept to this short list on purpose, a wider one makes j["key"] ambiguous
+    // note: a numeric one would make "str = json;" ambiguous, a number converts to char too
+    operator std::string() const;
+
+    template <typename T>
+    T value(const std::string & key, T def) const {
+        return contains(key) ? at(key).get<T>() : def;
+    }
+
+    std::string value(const std::string & key, const char * def) const;
+
+    // a JSON default needs no get<T>(), it is already the right type
+    common_json value(const std::string & key, const common_json & def) const {
+        return contains(key) ? at(key) : def;
+    }
+
+    void assign(const common_json_value & val);
+    void set(const common_json_item & item);
+    void push_back(const common_json_value & val);
+
+    // appends one object, e.g. push_back({{"a", 1}})
+    void push_back(std::initializer_list<common_json_item> items);
+
+    // 1 if the key is there, 0 if not
+    size_t count(const std::string & key) const;
+
+    // appends every value of another array; inserting an array into itself throws
+    void insert(const common_json & vals);
+
+    // a common_json goes through the copy assignment above, everything else becomes a value
+    template <typename T, typename std::enable_if<!std::is_same<typename std::decay<T>::type, common_json>::value, int>::type = 0>
+    common_json & operator=(T && val) {
+        assign(common_json_value(std::forward<T>(val)));
+        return *this;
+    }
+
+    std::string dump(int indent = -1) const;
+
+    // same as dump(), but bad UTF-8 gets replaced instead of throwing
+    std::string dump_safe(int indent = -1) const;
+
+    // walks an array by index, or an object in insertion order
+    // a plain value gives itself once, same as the backing library
+    class iterator {
+      public:
+        using iterator_category = std::forward_iterator_tag;
+        using value_type        = common_json;
+        using difference_type   = std::ptrdiff_t;
+        using pointer           = common_json *;
+        using reference         = common_json &;
+
+        iterator(common_json * node, size_t idx) : node(node), idx(idx) {}
+
+        common_json & operator*() const;
+        common_json & value()     const { return **this; }
+        std::string   key()       const;
+
+        iterator & operator++() {
+            idx++;
+            return *this;
+        }
+
+        bool operator!=(const iterator & other) const { return idx != other.idx; }
+        bool operator==(const iterator & other) const { return idx == other.idx; }
+
+      private:
+        common_json * node;
+        size_t        idx;
+    };
+
+    iterator begin() const;
+    iterator end()   const;
+
+    // allows: for (const auto & [key, val] : obj.items())
+    class items_view {
+      public:
+        // the members are public, so an entry also works with structured bindings
+        struct entry {
+            std::string   k;
+            common_json & v;
+
+            const std::string & key()   const { return k; }
+            common_json &       value() const { return v; }
+        };
+
+        items_view(common_json * node, size_t n) : node(node), n(n) {}
+
+        class iterator {
+          public:
+            iterator(common_json * node, size_t idx) : node(node), idx(idx) {}
+
+            entry operator*() const;
+
+            iterator & operator++() {
+                idx++;
+                return *this;
+            }
+
+            bool operator!=(const iterator & other) const { return idx != other.idx; }
+
+          private:
+            common_json * node;
+            size_t        idx;
+        };
+
+        iterator begin() const { return iterator(node, 0); }
+        iterator end()   const { return iterator(node, n); }
+
+      private:
+        common_json * node;
+        size_t        n;
+    };
+
+    items_view items() const;
+
+  private:
+    // a negative index must not turn into a huge size_t
+    static size_t to_idx(int idx) {
+        if (idx < 0) {
+            throw common_json_error("negative array index");
+        }
+        return (size_t) idx;
+    }
+
+    // the backing value is built here, json.cpp checks that it fits
+    // it cannot be a pointer: a value inside a tree would then not be a common_json
+    // at() could then only give back a copy instead of a real reference
+    alignas(8) unsigned char storage[32];
+};
+
+using common_json_entry = common_json::items_view::entry;
diff --git a/common/peg-parser.cpp b/common/peg-parser.cpp
index 4a4be7cf7..46fc29bf2 100644
--- a/common/peg-parser.cpp
+++ b/common/peg-parser.cpp
@@ -10,7 +10,6 @@
 #include <initializer_list>
 #include <map>
 #include <memory>
-#include <nlohmann/json.hpp>
 #include <regex>
 #include <set>
 #include <stdexcept>
@@ -1120,8 +1119,8 @@ common_peg_parser common_peg_parser_builder::chars(const std::string & classes,
     return wrap(arena_.add_parser(common_peg_chars_parser{classes, ranges, negated, min, max}));
 }
 
-common_peg_parser common_peg_parser_builder::schema(const common_peg_parser & p, const std::string & name, const nlohmann::ordered_json & schema, bool raw) {
-    return wrap(arena_.add_parser(common_peg_schema_parser{p.id(), name, std::make_shared<nlohmann::ordered_json>(schema), raw}));
+common_peg_parser common_peg_parser_builder::schema(const common_peg_parser & p, const std::string & name, const common_json & schema, bool raw) {
+    return wrap(arena_.add_parser(common_peg_schema_parser{p.id(), name, std::make_shared<common_json>(schema), raw}));
 }
 
 common_peg_parser common_peg_parser_builder::rule(const std::string & name, const common_peg_parser & p, bool trigger) {
@@ -1805,8 +1804,8 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
     }
 }
 
-static nlohmann::json serialize_parser_variant(const common_peg_parser_variant & variant) {
-    using json = nlohmann::json;
+static common_json serialize_parser_variant(const common_peg_parser_variant & variant) {
+    using json = common_json;
 
     return std::visit([](const auto & p) -> json {
         using T = std::decay_t<decltype(p)>;
@@ -1860,7 +1859,7 @@ static nlohmann::json serialize_parser_variant(const common_peg_parser_variant &
                 {"type", "schema"},
                 {"child", p.child},
                 {"name", p.name},
-                {"schema", p.schema ? *p.schema : nullptr},
+                {"schema", p.schema ? *p.schema : json(nullptr)},
                 {"raw", p.raw}
             };
         } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
@@ -1888,19 +1887,19 @@ static nlohmann::json serialize_parser_variant(const common_peg_parser_variant &
     }, variant);
 }
 
-nlohmann::json common_peg_arena::to_json() const {
-    auto parsers = nlohmann::json::array();
+common_json common_peg_arena::to_json() const {
+    auto parsers = common_json::array();
     for (const auto & parser : parsers_) {
         parsers.push_back(serialize_parser_variant(parser));
     }
-    return nlohmann::json{
+    return common_json{
         {"parsers", parsers},
         {"rules", rules_},
         {"root", root_}
     };
 }
 
-static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json & j) {
+static common_peg_parser_variant deserialize_parser_variant(const common_json & j) {
     if (!j.contains("type") || !j["type"].is_string()) {
         throw std::runtime_error("Parser variant JSON missing or invalid 'type' field");
     }
@@ -1969,9 +1968,9 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
         }
         common_peg_chars_parser parser;
         parser.pattern = j["pattern"];
-        parser.negated = j["negated"];
-        parser.min_count = j["min_count"];
-        parser.max_count = j["max_count"];
+        parser.negated = j["negated"].get<bool>();
+        parser.min_count = j["min_count"].get<int>();
+        parser.max_count = j["max_count"].get<int>();
         for (const auto & range_json : j["ranges"]) {
             if (!range_json.contains("start") || !range_json.contains("end")) {
                 throw std::runtime_error("char_range missing 'start' or 'end' field");
@@ -2007,7 +2006,7 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
         parser.child = j["child"].get<common_peg_parser_id>();
         parser.name = j["name"];
         if (!j["schema"].is_null()) {
-            parser.schema = std::make_shared<nlohmann::ordered_json>(j["schema"]);
+            parser.schema = std::make_shared<common_json>(j["schema"]);
         }
         parser.raw = j["raw"].get<bool>();
         return parser;
@@ -2069,7 +2068,7 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
     throw std::runtime_error("Unknown parser type: " + type);
 }
 
-common_peg_arena common_peg_arena::from_json(const nlohmann::json & j) {
+common_peg_arena common_peg_arena::from_json(const common_json & j) {
     if (!j.contains("parsers") || !j["parsers"].is_array()) {
         throw std::runtime_error("JSON missing or invalid 'parsers' array");
     }
@@ -2109,7 +2108,7 @@ std::string common_peg_arena::save() const {
 }
 
 void common_peg_arena::load(const std::string & data) {
-    *this = from_json(nlohmann::json::parse(data));
+    *this = from_json(common_json::parse(data));
 }
 
 common_peg_arena build_peg_parser(const std::function<common_peg_parser(common_peg_parser_builder & builder)> & fn) {
diff --git a/common/peg-parser.h b/common/peg-parser.h
index c198499dd..ab095cc7d 100644
--- a/common/peg-parser.h
+++ b/common/peg-parser.h
@@ -1,6 +1,6 @@
 #pragma once
 
-#include <nlohmann/json_fwd.hpp>
+#include "json.h"
 
 #include <memory>
 #include <set>
@@ -245,7 +245,7 @@ struct common_peg_until_parser {
 struct common_peg_schema_parser {
     common_peg_parser_id child;
     std::string name;
-    std::shared_ptr<nlohmann::ordered_json> schema;
+    std::shared_ptr<common_json> schema;
 
     // Indicates if the GBNF should accept a raw string that matches the schema.
     bool raw;
@@ -332,8 +332,8 @@ class common_peg_arena {
 
     std::string dump(common_peg_parser_id id) const;
 
-    nlohmann::json to_json() const;
-    static common_peg_arena from_json(const nlohmann::json & j);
+    common_json to_json() const;
+    static common_peg_arena from_json(const common_json & j);
 
     std::string save() const;
     void load(const std::string & data);
@@ -490,7 +490,7 @@ class common_peg_parser_builder {
 
     // Wraps a parser with JSON schema metadata for grammar generation.
     // Used internally to convert JSON schemas to GBNF grammar rules.
-    common_peg_parser schema(const common_peg_parser & p, const std::string & name, const nlohmann::ordered_json & schema, bool raw = false);
+    common_peg_parser schema(const common_peg_parser & p, const std::string & name, const common_json & schema, bool raw = false);
 
     // Creates a named rule, stores it in the grammar, and returns a ref.
     // If trigger=true, marks this rule as an entry point for lazy grammar generation.
diff --git a/conversion/glm.py b/conversion/glm.py
index abd7f279f..7544f850c 100644
--- a/conversion/glm.py
+++ b/conversion/glm.py
@@ -112,12 +112,38 @@ class GlmOCRModel(Glm4Model):
 @ModelBase.example("zai-org/GLM-4.5-Air")
 class Glm4MoeModel(TextModel):
     model_arch = gguf.MODEL_ARCH.GLM4_MOE
+    supports_mtp_export = True
+    _n_main_layers: int | None = None
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        # GLM4_MOE has num_hidden_layers + 1 actual layers (including NextN layer)
-        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
-        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+        if not self.no_mtp:
+            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
+            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
+        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
+        type(self)._n_main_layers = hparams.get(key)
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+
+        assert cls._n_main_layers is not None
+        is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers
+
+        if is_mtp and cls.no_mtp:
+            return None
+        if cls.mtp_only and not is_mtp and name not in (
+            "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+        ):
+            return None
+
+        return name, gen
 
     def set_vocab(self):
         return self._set_vocab_glm()
@@ -153,10 +179,22 @@ class Glm4MoeModel(TextModel):
         if (norm_topk_prob := self.hparams.get("norm_topk_prob")) is not None:
             self.gguf_writer.add_expert_weights_norm(norm_topk_prob)
 
-        # NextN/MTP prediction layers
-        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+        if not self.no_mtp and (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
             self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
 
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
     _experts: list[dict[str, Tensor]] | None = None
 
     # note: unlike GLM4V non-MoE, we don't need to permute Q/K here since GLM4V_MOE uses Neox ordering already
@@ -348,6 +386,7 @@ class GlmMoeDsaModel(DeepseekV2Model):
 @ModelBase.example("upstage/Solar-Open-100B")
 class SolarOpenModel(Glm4MoeModel):
     model_arch = gguf.MODEL_ARCH.GLM4_MOE
+    supports_mtp_export = False
 
     def set_vocab(self):
         from transformers import AutoTokenizer
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 40c6e3bd8..70e38e4e7 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -1742,6 +1742,19 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   n_past);
 
+    GGML_API struct ggml_tensor * ggml_clamp(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 min,
+            float                 max);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_clamp_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 min,
+            float                 max);
+
     GGML_API struct ggml_tensor * ggml_soft_max(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -2008,14 +2021,6 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   n_offs);
 
-    // clamp
-    // in-place, returns view(a)
-    GGML_API struct ggml_tensor * ggml_clamp(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            float                 min,
-            float                 max);
-
     // im2col
     // converts data into a format that effectively results in a convolution when combined with matrix multiplication
     GGML_API struct ggml_tensor * ggml_im2col(
diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c
index 3bda9abbe..a71838eaf 100644
--- a/ggml/src/ggml-alloc.c
+++ b/ggml/src/ggml-alloc.c
@@ -40,6 +40,7 @@ bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_SILU_BACK:
         case GGML_OP_RMS_NORM:
         case GGML_OP_RMS_NORM_BACK:
+        case GGML_OP_CLAMP:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
             return true;
diff --git a/ggml/src/ggml-backend-meta.cpp b/ggml/src/ggml-backend-meta.cpp
index 7654ea1f3..fe58ea3bb 100644
--- a/ggml/src/ggml-backend-meta.cpp
+++ b/ggml/src/ggml-backend-meta.cpp
@@ -592,7 +592,18 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
             GGML_ASSERT(split_states_equal(src_ss[0], src_ss[1]));
             return {assume_sync ? GGML_BACKEND_SPLIT_AXIS_MIRRORED : GGML_BACKEND_SPLIT_AXIS_PARTIAL, {0}, {1}, 1};
         }
-        GGML_ABORT("fatal error");
+        if (src_ss[0].axis == src_ss[1].axis && src_ss[0].axis >= GGML_BACKEND_SPLIT_AXIS_2 &&
+                src_ss[0].axis < GGML_MAX_DIMS) {
+            GGML_ASSERT(split_states_equal(src_ss[0], src_ss[1]));
+            return src_ss[0];
+        }
+        // batched matmul with the batches split across devices and a replicated activation
+        if (src_ss[0].axis >= GGML_BACKEND_SPLIT_AXIS_2 && src_ss[0].axis < GGML_MAX_DIMS &&
+                src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+            return src_ss[0];
+        }
+        GGML_ABORT("unsupported mul_mat split states: node=%s src0=%s axis=%d src1=%s axis=%d",
+            tensor->name, tensor->src[0]->name, (int) src_ss[0].axis, tensor->src[1]->name, (int) src_ss[1].axis);
         //return {GGML_BACKEND_SPLIT_AXIS_UNKNOWN, {0}, {1}, 1};
     };
 
@@ -602,27 +613,40 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
             case GGML_BACKEND_SPLIT_AXIS_1:
             case GGML_BACKEND_SPLIT_AXIS_2:
             case GGML_BACKEND_SPLIT_AXIS_3: {
-                GGML_ASSERT(src_ss[0].n_segments == 1);
-                if (src_ss[0].axis == ggml_n_dims(tensor->src[0]) - 1 && src_ss[0].nr[0] == 1) {
-                    return {ggml_backend_meta_split_axis(ggml_n_dims(tensor) - 1), {0}, {1}, 1};
-                }
-                int64_t base_ne_in = tensor->src[0]->ne[0];
-                for (int dim = 1; dim <= src_ss[0].axis; dim++) {
+                int64_t base_ne_in = 1;
+                for (int dim = 0; dim <= src_ss[0].axis; dim++) {
                     base_ne_in *= tensor->src[0]->ne[dim];
                 }
-                base_ne_in /= src_ss[0].nr[0];
+                if (src_ss[0].n_segments == 1) {
+                    base_ne_in /= src_ss[0].nr[0];
+                    if (src_ss[0].axis == ggml_n_dims(tensor->src[0]) - 1 && src_ss[0].nr[0] == 1) {
+                        return {ggml_backend_meta_split_axis(ggml_n_dims(tensor) - 1), {0}, {1}, 1};
+                    }
+                    if (src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_0 && tensor->ne[0] == tensor->src[0]->ne[0] &&
+                            tensor->ne[1] == 1 && src_ss[0].nr[0] == 1) {
+                        bool complete_rows = true;
+                        for (size_t j = 0; j < n_bufs; j++) {
+                            const int64_t ne = src_ss[0].ne[j];
+                            complete_rows = complete_rows && (ne == 0 || ne == tensor->src[0]->ne[0]);
+                        }
+                        if (complete_rows) {
+                            // Move a complete dim-0 split to the following singleton dimension.
+                            return {GGML_BACKEND_SPLIT_AXIS_1, {0}, {1}, 1};
+                        }
+                    }
+                }
+                // Reshape outputs use one segment; split-state propagation merges source segments.
                 int64_t base_ne_out = 1;
                 for (int dim = 0; dim < GGML_MAX_DIMS; dim++) {
-                    const int64_t base_ne_out_next = base_ne_out *= tensor->ne[dim];
-                    if (base_ne_out_next % base_ne_in == 0) {
-                        return {ggml_backend_meta_split_axis(dim), {0}, {uint32_t(base_ne_out_next/base_ne_in)}, 1};
+                    base_ne_out *= tensor->ne[dim];
+                    if (base_ne_out % base_ne_in == 0) {
+                        return {ggml_backend_meta_split_axis(dim), {0}, {uint32_t(base_ne_out/base_ne_in)}, 1};
                     }
-                    if (base_ne_out_next > base_ne_in) {
+                    if (base_ne_out > base_ne_in) {
                         GGML_ASSERT(src_ss[0].n_segments == 1);
                         GGML_ASSERT(src_ss[0].nr[0]      == 1);
                         return {ggml_backend_meta_split_axis(dim), {0}, {1}, 1};
                     }
-                    base_ne_out = base_ne_out_next;
                 }
                 GGML_ABORT("shape mismatch for %s", ggml_op_name(tensor->op));
             }
@@ -747,14 +771,33 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
     };
 
     auto handle_flash_attn_ext = [&](const std::vector<ggml_backend_meta_split_state> & src_ss) -> ggml_backend_meta_split_state {
-        GGML_ASSERT(                             src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_2);
-        GGML_ASSERT(                             src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_2);
-        GGML_ASSERT(                             src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_2);
-        GGML_ASSERT(tensor->src[4] == nullptr || src_ss[3].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+        GGML_ASSERT(tensor->src[3] == nullptr || src_ss[3].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+
+        if (src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+            GGML_ASSERT(src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+            GGML_ASSERT(src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+            GGML_ASSERT(tensor->src[4] == nullptr || src_ss[4].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+            return {GGML_BACKEND_SPLIT_AXIS_MIRRORED, {0}, {1}, 1};
+        }
+
+        GGML_ASSERT(src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_2);
+        const bool kv_split = src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_2 &&
+                src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_2;
+        const bool kv_mirrored = src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED &&
+                src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED;
+        GGML_ASSERT(kv_split || kv_mirrored);
         GGML_ASSERT(tensor->src[4] == nullptr || src_ss[4].axis == GGML_BACKEND_SPLIT_AXIS_0);
         return {GGML_BACKEND_SPLIT_AXIS_1, {0}, {1}, 1};
     };
 
+    auto handle_lightning_indexer = [&](
+            const std::vector<ggml_backend_meta_split_state> & src_ss) -> ggml_backend_meta_split_state {
+        for (size_t i = 0; i < 4; i++) {
+            GGML_ASSERT(src_ss[i].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+        }
+        return {GGML_BACKEND_SPLIT_AXIS_MIRRORED, {0}, {1}, 1};
+    };
+
     auto handle_ssm_conv = [&](const std::vector<ggml_backend_meta_split_state> & src_ss) -> ggml_backend_meta_split_state {
         if (src_ss[0].axis == src_ss[1].axis) {
             if (src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_0) {
@@ -792,7 +835,7 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
             ggml_backend_dev_t dev = ggml_backend_buft_get_device(ggml_backend_buffer_get_type(tensor->buffer));
             const ggml_backend_meta_device_context * dev_ctx = (const ggml_backend_meta_device_context *) dev->context;
             ggml_backend_meta_split_state ret = dev_ctx->get_split_state(tensor, dev_ctx->get_split_state_ud);
-            if (ret.axis >= 0 && ret.axis <= GGML_MAX_DIMS) {
+            if (ret.axis >= 0 && ret.axis < GGML_MAX_DIMS) {
                 const int64_t granularity = ret.axis == GGML_BACKEND_SPLIT_AXIS_0 ? ggml_blck_size(tensor->type) : 1;
                 int64_t ne_sum = 0;
                 for (size_t s = 0; s < ret.n_segments; s++) {
@@ -802,6 +845,9 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
                     }
                 }
                 GGML_ASSERT(ne_sum == tensor->ne[ret.axis]);
+            } else if (ret.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
+                GGML_ASSERT(ret.n_segments == 1);
+                GGML_ASSERT(ret.nr[0] == 1);
             }
             return ret;
         }
@@ -922,7 +968,7 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
                 split_state = handle_rope(src_ss);
             } break;
             case GGML_OP_ROPE_BACK: {
-                split_state = handle_generic(src_ss, /*scalar_only =*/ true);
+                split_state = handle_rope(src_ss);
             } break;
             case GGML_OP_CLAMP: {
                 split_state = handle_generic(src_ss, /*scalar_only =*/ false);
@@ -986,6 +1032,9 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
             case GGML_OP_GATED_DELTA_NET: {
                 split_state = handle_gated_delta_net(src_ss);
             } break;
+            case GGML_OP_LIGHTNING_INDEXER: {
+                split_state = handle_lightning_indexer(src_ss);
+            } break;
             case GGML_OP_DSV4_HC_COMB:
             case GGML_OP_DSV4_HC_PRE:
             case GGML_OP_DSV4_HC_POST: {
@@ -1070,13 +1119,14 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
         if (buf_ctx->debug > 0) {
             std::string srcs_info;
             for (size_t i = 0; i < GGML_MAX_SRC; i++) {
-                if (tensor->src[i] == nullptr) {
+                if (tensor->src[i] == nullptr || tensor->src[i] == tensor) {
                     continue;
                 }
                 if (!srcs_info.empty()) {
                     srcs_info += ", ";
                 }
-                const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(tensor->src[0], true);
+                const ggml_backend_meta_split_state split_state =
+                        ggml_backend_meta_get_split_state(tensor->src[i], true);
                 GGML_ASSERT(split_state.n_segments == 1);
                 const char * axis_name = ggml_backend_meta_split_axis_name(split_state.axis);
                 std::string ne_info;
@@ -1255,6 +1305,108 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor(ggml_backend_buffer
     return ggml_backend_meta_buffer_init_tensor_impl(buf_ctx->get_simple_tensor_container(tensor), tensor);
 }
 
+static void ggml_backend_meta_buffer_memset_tensor(
+        ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    const size_t n_bufs = ggml_backend_meta_buffer_n_bufs(buffer);
+    const ggml_backend_meta_split_state split_state =
+            ggml_backend_meta_get_split_state(tensor, /*assume_sync =*/ false);
+    GGML_ASSERT(ggml_is_contiguous(tensor) || split_state.axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+
+    if (split_state.n_segments != 1 || split_state.nr[0] != 1) {
+        GGML_ASSERT(split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS);
+        GGML_ASSERT(split_state.nr[0] != 0);
+        GGML_ASSERT(tensor->ne[3] == 1);
+
+        std::vector<size_t> simple_offsets(n_bufs, 0);
+        if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_0) {
+            GGML_ASSERT(tensor->ne[2] == 1);
+
+            const size_t row_stride = tensor->nb[1];
+            GGML_ASSERT(offset % row_stride == 0);
+            GGML_ASSERT(size   % row_stride == 0);
+            const int64_t row_start = offset / row_stride;
+            const int64_t row_count = size   / row_stride;
+            GGML_ASSERT(row_start + row_count <= tensor->ne[1]);
+
+            const int64_t blck_size = ggml_blck_size(tensor->type);
+            for (size_t s = 0; s < split_state.n_segments; s++) {
+                for (size_t r = 0; r < split_state.nr[s]; r++) {
+                    for (size_t j = 0; j < n_bufs; j++) {
+                        ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
+                        GGML_ASSERT(split_state.ne[s*n_bufs + j] % blck_size == 0);
+                        const size_t nbytes = split_state.ne[s*n_bufs + j]/blck_size * tensor->nb[0];
+                        for (int64_t row = 0; row < row_count; row++) {
+                            ggml_backend_tensor_memset(simple_tensor, value,
+                                    simple_offsets[j] + (row_start + row)*simple_tensor->nb[1], nbytes);
+                        }
+                        simple_offsets[j] += nbytes;
+                    }
+                }
+            }
+            return;
+        }
+
+        GGML_ASSERT(split_state.axis == GGML_BACKEND_SPLIT_AXIS_1);
+
+        const size_t row_stride = tensor->nb[2];
+        GGML_ASSERT(offset % row_stride == 0);
+        GGML_ASSERT(size   % row_stride == 0);
+        const int64_t row_start = offset / row_stride;
+        const int64_t row_count = size   / row_stride;
+        GGML_ASSERT(row_start + row_count <= tensor->ne[2]);
+
+        for (size_t s = 0; s < split_state.n_segments; s++) {
+            for (size_t r = 0; r < split_state.nr[s]; r++) {
+                for (size_t j = 0; j < n_bufs; j++) {
+                    ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
+                    const size_t nbytes = split_state.ne[s*n_bufs + j] * tensor->nb[1];
+                    for (int64_t row = 0; row < row_count; row++) {
+                        ggml_backend_tensor_memset(simple_tensor, value,
+                                simple_offsets[j] + (row_start + row)*simple_tensor->nb[2], nbytes);
+                    }
+                    simple_offsets[j] += nbytes;
+                }
+            }
+        }
+        return;
+    }
+
+    switch (split_state.axis) {
+        case GGML_BACKEND_SPLIT_AXIS_0:
+        case GGML_BACKEND_SPLIT_AXIS_1:
+        case GGML_BACKEND_SPLIT_AXIS_2: {
+            const size_t chunk_size_full = tensor->nb[split_state.axis + 1];
+            GGML_ASSERT(offset % chunk_size_full == 0);
+            GGML_ASSERT(size   % chunk_size_full == 0);
+            const int64_t i_start =  offset        / chunk_size_full;
+            const int64_t i_stop  = (offset + size) / chunk_size_full;
+            for (size_t j = 0; j < n_bufs; j++) {
+                ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
+                const size_t chunk_size = simple_tensor->nb[split_state.axis + 1];
+                if (chunk_size == 0) {
+                    continue;
+                }
+                for (int64_t i = i_start; i < i_stop; i++) {
+                    ggml_backend_tensor_memset(simple_tensor, value, i*chunk_size, chunk_size);
+                }
+            }
+        } break;
+        case GGML_BACKEND_SPLIT_AXIS_PARTIAL: {
+            GGML_ASSERT(value == 0);
+            [[fallthrough]];
+        }
+        case GGML_BACKEND_SPLIT_AXIS_MIRRORED: {
+            for (size_t j = 0; j < n_bufs; j++) {
+                ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
+                ggml_backend_tensor_memset(simple_tensor, value, offset, size);
+            }
+        } break;
+        default: {
+            GGML_ABORT("fatal error");
+        }
+    }
+}
+
 static void ggml_backend_meta_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     const size_t n_bufs = ggml_backend_meta_buffer_n_bufs(buffer);
     const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(tensor, /*assume_sync =*/ false);
@@ -1352,15 +1504,29 @@ static void ggml_backend_meta_buffer_set_tensor(ggml_backend_buffer_t buffer, gg
         } break;
         case GGML_BACKEND_SPLIT_AXIS_PARTIAL: {
             GGML_ASSERT(tensor->type == GGML_TYPE_F32);
-            const int64_t ne = ggml_nelements(tensor);
-            std::vector<float> tmp;
-            tmp.reserve(ne);
-            for (int64_t i = 0; i < ne; i++) {
-                tmp.push_back(((const float *) data)[i] / n_bufs);
+            GGML_ASSERT(offset % sizeof(float) == 0);
+            GGML_ASSERT(size   % sizeof(float) == 0);
+            const size_t n_values = size / sizeof(float);
+            size_t n_contributors = 0;
+            for (size_t j = 0; j < n_bufs; j++) {
+                n_contributors += split_state.ne[j] != 0;
+            }
+            const bool has_contributor_mask = n_contributors != 0;
+            if (!has_contributor_mask) {
+                n_contributors = n_bufs;
+            }
+            std::vector<float> tmp(n_values);
+            for (size_t i = 0; i < n_values; i++) {
+                tmp[i] = ((const float *) data)[i] / n_contributors;
+            }
+            std::vector<float> zero;
+            if (has_contributor_mask) {
+                zero.resize(n_values, 0.0f);
             }
             for (size_t j = 0; j < n_bufs; j++) {
                 ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
-                ggml_backend_tensor_set(simple_tensor, tmp.data(), offset, size);
+                const float * partial = has_contributor_mask && split_state.ne[j] == 0 ? zero.data() : tmp.data();
+                ggml_backend_tensor_set(simple_tensor, partial, offset, size);
             }
         } break;
         default: {
@@ -1488,7 +1654,7 @@ static const ggml_backend_buffer_i ggml_backend_meta_buffer_iface = {
     /* .free_buffer     = */ ggml_backend_meta_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_meta_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_meta_buffer_init_tensor,
-    /* .memset_tensor   = */ nullptr, // TODO implement
+    /* .memset_tensor   = */ ggml_backend_meta_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_meta_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_meta_buffer_get_tensor,
     /* .set_tensor_2d   = */ nullptr,
@@ -1841,7 +2007,7 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
 
         {
             // For MoE models it may make sense to delay the AllReduce in order to reduce I/O:
-            auto get_i_delayed = [&](const int i) -> int {
+            auto get_i_delayed_branch = [&](const int i) -> int {
                 int id = i; // i_delayed
                 int idr = i; // i_delayed return, last safe return value
 
@@ -1941,6 +2107,62 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
                 return idr;
             };
 
+            // AllReduce(a) + AllReduce(b) == AllReduce(a + b) for independent partial branches.
+            auto get_i_delayed = [&](const int i) -> int {
+                const int i_delayed = get_i_delayed_branch(i);
+                ggml_tensor * node = cgraph->nodes[i_delayed];
+
+                if (ggml_node_get_use_count(cgraph, i_delayed) != 1) {
+                    return i_delayed;
+                }
+
+                for (int id = i_delayed + 1; id < cgraph->n_nodes; id++) {
+                    ggml_tensor * next = cgraph->nodes[id];
+                    if (next->view_src == node) {
+                        return i_delayed;
+                    }
+                    for (int s = 0; s < GGML_MAX_SRC; s++) {
+                        if (next->src[s] == node) {
+                            return i_delayed;
+                        }
+                    }
+
+                    if (next->view_src != nullptr && next->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(next->view_src->buffer)) {
+                        continue;
+                    }
+                    if (ggml_backend_meta_get_split_state(next, false).axis != GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
+                        continue;
+                    }
+
+                    const int i_other = id;
+                    const int i_other_delayed = get_i_delayed_branch(i_other);
+                    ggml_tensor * other = cgraph->nodes[i_other_delayed];
+                    if (ggml_node_get_use_count(cgraph, i_other_delayed) != 1 || i_other_delayed + 1 >= cgraph->n_nodes) {
+                        return i_delayed;
+                    }
+
+                    ggml_tensor * sum = cgraph->nodes[i_other_delayed + 1];
+                    if (sum->op != GGML_OP_ADD ||
+                            !ggml_are_same_shape(node, other) || node->type != other->type || sum->type != node->type ||
+                            !((sum->src[0] == node && sum->src[1] == other) ||
+                              (sum->src[0] == other && sum->src[1] == node)) ||
+                            ggml_backend_meta_get_split_state(sum, false).axis != GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                        return i_delayed;
+                    }
+
+                    for (size_t j = 0; j < n_backends; j++) {
+                        auto & bcj = backend_ctx->backend_configs[j];
+                        const bool compute       = bcj.nodes[i]->flags       & GGML_TENSOR_FLAG_COMPUTE;
+                        const bool compute_other = bcj.nodes[i_other]->flags & GGML_TENSOR_FLAG_COMPUTE;
+                        if (compute != compute_other) {
+                            return i_delayed;
+                        }
+                    }
+                    return i_other_delayed + 1;
+                }
+                return i_delayed;
+            };
+
             int i_start = 0;
             for (int i = 0; i < cgraph->n_nodes; i++) {
                 ggml_tensor * node = cgraph->nodes[i];
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index e3fa8304a..5d7768c5d 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -40,6 +40,7 @@ bool g_mul_mat_q = true;
 #include "ggml-cuda/out-prod.cuh"
 #include "ggml-cuda/pad.cuh"
 #include "ggml-cuda/pool2d.cuh"
+#include "ggml-cuda/pool1d.cuh"
 #include "ggml-cuda/quantize.cuh"
 #include "ggml-cuda/rope.cuh"
 #include "ggml-cuda/roll.cuh"
@@ -2330,6 +2331,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_POOL_2D:
             ggml_cuda_op_pool2d(ctx, dst);
             break;
+        case GGML_OP_POOL_1D:
+            ggml_cuda_op_pool1d(ctx, dst);
+            break;
         case GGML_OP_SUM:
             ggml_cuda_op_sum(ctx, dst);
             break;
@@ -4620,8 +4624,8 @@ static std::string ggml_cuda_device_description(int device) {
     const ggml_cuda_device_info & info = ggml_cuda_info();
     std::string description = prop.name;
     if (info.device_count > info.physical_device_count) {
-        description += " (physical device " + std::to_string(info.devices[device].physical_device) +
-                       ", virtual device " + std::to_string(info.devices[device].virtual_index) + ")";
+        description += " (dev p" + std::to_string(info.devices[device].physical_device) +
+                       "/v" + std::to_string(info.devices[device].virtual_index) + ")";
     }
     return description;
 }
@@ -5258,6 +5262,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONV_2D_DW:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_CONV_TRANSPOSE_2D:
+        case GGML_OP_POOL_1D:
         case GGML_OP_POOL_2D:
             return true;
         case GGML_OP_ACC:
diff --git a/ggml/src/ggml-cuda/pool1d.cu b/ggml/src/ggml-cuda/pool1d.cu
new file mode 100644
index 000000000..ac6fb0cbd
--- /dev/null
+++ b/ggml/src/ggml-cuda/pool1d.cu
@@ -0,0 +1,85 @@
+#include "pool1d.cuh"
+
+static __global__ void pool1d_nchw_kernel(
+        const int iw, const int ow,
+        const int kw, const int sw, const int pw,
+        const int parallel_elements,
+        const float * src, float * dst, const enum ggml_op_pool op) {
+    const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (idx >= parallel_elements) {
+        return;
+    }
+
+    const int nc = idx / ow;
+    const int cur_ow = idx % ow;
+
+    const float * i_ptr = src + nc * iw;
+    float * o_ptr = dst + nc * ow;
+
+    const int start = cur_ow * sw - pw;
+    const int b = max(0, start);
+    const int e = min(iw, start + kw);
+
+    float res;
+    switch (op) {
+        case GGML_OP_POOL_AVG: res = 0.0f;     break;
+        case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
+        default: return;
+    }
+
+    int count = 0;
+    for (int i = b; i < e; i++) {
+#if __CUDA_ARCH__ >= 350
+        float cur = __ldg(i_ptr + i);
+#else
+        float cur = i_ptr[i];
+#endif
+        switch (op) {
+            case GGML_OP_POOL_AVG: res += cur;                break;
+            case GGML_OP_POOL_MAX: res = max(res, cur);       break;
+            default: break;
+        }
+        count++;
+    }
+
+    if (op == GGML_OP_POOL_AVG) {
+        res = (count > 0) ? (res / count) : 0.0f;
+    }
+
+    o_ptr[cur_ow] = res;
+}
+
+static void pool1d_nchw_kernel_f32_f32_cuda(
+        const int iw, const int ow,
+        const int kw, const int sw, const int pw,
+        const int parallel_elements,
+        const float * src, float * dst, const enum ggml_op_pool op,
+        cudaStream_t stream) {
+    const int num_blocks = (parallel_elements + CUDA_POOL1D_BLOCK_SIZE - 1) / CUDA_POOL1D_BLOCK_SIZE;
+    dim3 block_nums(num_blocks);
+    pool1d_nchw_kernel<<<block_nums, CUDA_POOL1D_BLOCK_SIZE, 0, stream>>>(iw, ow, kw, sw, pw, parallel_elements, src, dst, op);
+}
+
+void ggml_cuda_op_pool1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
+    const int k0 = opts[1];
+    const int s0 = opts[2];
+    const int p0 = opts[3];
+
+    const int64_t IW = src0->ne[0];
+    const int64_t OW = dst->ne[0];
+    const int64_t nr = ggml_nrows(src0);
+
+    const int parallel_elements = (int)(nr * OW);
+
+    pool1d_nchw_kernel_f32_f32_cuda(IW, OW, k0, s0, p0, parallel_elements, src0_d, dst_d, op, stream);
+}
diff --git a/ggml/src/ggml-cuda/pool1d.cuh b/ggml/src/ggml-cuda/pool1d.cuh
new file mode 100644
index 000000000..c79461dd8
--- /dev/null
+++ b/ggml/src/ggml-cuda/pool1d.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_POOL1D_BLOCK_SIZE 256
+
+void ggml_cuda_op_pool1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-metal/ggml-metal-device.cpp b/ggml/src/ggml-metal/ggml-metal-device.cpp
index 52043696e..24caad7ff 100644
--- a/ggml/src/ggml-metal/ggml-metal-device.cpp
+++ b/ggml/src/ggml-metal/ggml-metal-device.cpp
@@ -17,10 +17,10 @@ struct ggml_metal_device_deleter {
 
 typedef std::unique_ptr<ggml_metal_device, ggml_metal_device_deleter> ggml_metal_device_ptr;
 
-ggml_metal_device_t ggml_metal_device_get(int device) {
+ggml_metal_device_t ggml_metal_device_get(int device, int n_devices) {
     static std::vector<ggml_metal_device_ptr> devs;
 
-    devs.emplace_back(ggml_metal_device_init(device));
+    devs.emplace_back(ggml_metal_device_init(device, n_devices));
 
     return devs.back().get();
 }
diff --git a/ggml/src/ggml-metal/ggml-metal-device.h b/ggml/src/ggml-metal/ggml-metal-device.h
index b7d466058..5f5410a03 100644
--- a/ggml/src/ggml-metal/ggml-metal-device.h
+++ b/ggml/src/ggml-metal/ggml-metal-device.h
@@ -259,6 +259,8 @@ enum ggml_metal_device_id {
 
 struct ggml_metal_device_props {
     int device;
+    int device_phys;
+    int device_virt;
     char name[128];
     char desc[128];
 
@@ -286,10 +288,10 @@ typedef struct ggml_metal_event * ggml_metal_event_t;
 void ggml_metal_event_encode_signal(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf);
 void ggml_metal_event_encode_wait  (ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf);
 
-ggml_metal_device_t ggml_metal_device_init(int device);
+ggml_metal_device_t ggml_metal_device_init(int device, int n_devices);
 void ggml_metal_device_free(ggml_metal_device_t dev);
 
-ggml_metal_device_t ggml_metal_device_get(int device);
+ggml_metal_device_t ggml_metal_device_get(int device, int n_devices);
 
 void * ggml_metal_device_get_obj  (ggml_metal_device_t dev); // id<MTLDevice>
 void * ggml_metal_device_get_queue(ggml_metal_device_t dev); // id<MTLCommandQueue>
diff --git a/ggml/src/ggml-metal/ggml-metal-device.m b/ggml/src/ggml-metal/ggml-metal-device.m
index bccc86c9e..35ffe0813 100644
--- a/ggml/src/ggml-metal/ggml-metal-device.m
+++ b/ggml/src/ggml-metal/ggml-metal-device.m
@@ -717,7 +717,7 @@ static enum ggml_metal_device_id ggml_metal_device_id_parse(const char * name) {
     return GGML_METAL_DEVICE_GENERIC;
 }
 
-ggml_metal_device_t ggml_metal_device_init(int device) {
+ggml_metal_device_t ggml_metal_device_init(int device, int n_devices) {
     ggml_metal_device_t dev = calloc(1, sizeof(struct ggml_metal_device));
 
     assert(dev != NULL);
@@ -734,6 +734,12 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
             dev->addr_virt = 0x000000400ULL;
 
             dev->props.device = device;
+
+            // the Metal backend uses the system default device as the single physical device;
+            // additional (virtual) devices are emulated on top of it via GGML_METAL_DEVICES
+            dev->props.device_phys = 0;
+            dev->props.device_virt = device;
+
             dev->props.has_simdgroup_reduction  = [dev->mtl_device supportsFamily:MTLGPUFamilyApple7];
             dev->props.has_simdgroup_reduction |= [dev->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
 
@@ -897,7 +903,13 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
             }
 
             snprintf(dev->props.name, sizeof(dev->props.name), "%s%d", "MTL", device);
-            snprintf(dev->props.desc, sizeof(dev->props.desc), "%s", [[dev->mtl_device name] UTF8String]);
+            const char * gpu_name = [[dev->mtl_device name] UTF8String];
+            if (n_devices > 1) {
+                snprintf(dev->props.desc, sizeof(dev->props.desc), "%s (dev p%d/v%d)",
+                         gpu_name, dev->props.device_phys, dev->props.device_virt);
+            } else {
+                snprintf(dev->props.desc, sizeof(dev->props.desc), "%s", gpu_name);
+            }
 
             dev->library = ggml_metal_library_init(dev);
             if (!dev->library) {
diff --git a/ggml/src/ggml-metal/ggml-metal.cpp b/ggml/src/ggml-metal/ggml-metal.cpp
index 0e8d409e0..31aa61e32 100644
--- a/ggml/src/ggml-metal/ggml-metal.cpp
+++ b/ggml/src/ggml-metal/ggml-metal.cpp
@@ -891,7 +891,7 @@ static ggml_backend_dev_t ggml_backend_metal_device_init(ggml_backend_reg_t reg,
     return new ggml_backend_device {
         /* .iface   = */ ggml_backend_metal_device_i,
         /* .reg     = */ reg,
-        /* .context = */ ggml_metal_device_get(device),
+        /* .context = */ ggml_metal_device_get(device, g_devices),
     };
 }
 
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index cc44cdd79..e6b364fa7 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -961,6 +961,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_diag[2];
     vk_pipeline pipeline_clamp[2];
     vk_pipeline pipeline_pad_f32;
+    vk_pipeline pipeline_pad_reflect_1d_f32;
     vk_pipeline pipeline_roll_f32;
     vk_pipeline pipeline_repeat_i32, pipeline_repeat_back_f32;
     vk_pipeline pipeline_repeat_i16;
@@ -5636,6 +5637,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_diag[1], "diag_f16", diag_f16_len, diag_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_pad_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_pad_reflect_1d_f32, "pad_reflect_1d_f32", pad_reflect_1d_f32_len, pad_reflect_1d_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_roll_f32, "roll_f32", roll_f32_len, roll_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -11369,6 +11371,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_pad_f32;
         }
         return nullptr;
+    case GGML_OP_PAD_REFLECT_1D:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_pad_reflect_1d_f32;
+        }
+        return nullptr;
     case GGML_OP_ROLL:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_roll_f32;
@@ -12272,6 +12279,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_CLAMP:
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_PAD:
+    case GGML_OP_PAD_REFLECT_1D:
     case GGML_OP_ROLL:
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
@@ -13144,6 +13152,17 @@ static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const
     ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD, std::move(p));
 }
 
+static void ggml_vk_pad_reflect_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    const uint32_t p0 = (uint32_t)dst->op_params[0];
+    const uint32_t p1 = (uint32_t)dst->op_params[1];
+
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
+    memcpy(&p.param1, &p0, sizeof(float));
+    memcpy(&p.param2, &p1, sizeof(float));
+
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD_REFLECT_1D, std::move(p));
+}
+
 static void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const int32_t s0 = ggml_get_op_params_i32(dst, 0);
     const int32_t s1 = ggml_get_op_params_i32(dst, 1);
@@ -15553,6 +15572,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_PAD:
         ggml_vk_pad(ctx, compute_ctx, src0, node);
 
+        break;
+    case GGML_OP_PAD_REFLECT_1D:
+        ggml_vk_pad_reflect_1d(ctx, compute_ctx, src0, node);
+
         break;
     case GGML_OP_ROLL:
         ggml_vk_roll(ctx, compute_ctx, src0, node);
@@ -18479,6 +18502,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SCALE:
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_PAD:
+        case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ROLL:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_DIAG_MASK_INF:
@@ -19261,6 +19285,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         } else if (tensor->op == GGML_OP_PAD) {
             tensor_clone = ggml_pad_ext(ggml_ctx, src_clone[0], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3],
                                                                 tensor->op_params[4], tensor->op_params[5], tensor->op_params[6], tensor->op_params[7]);
+        } else if (tensor->op == GGML_OP_PAD_REFLECT_1D) {
+            tensor_clone = ggml_pad_reflect_1d(ggml_ctx, src_clone[0], tensor->op_params[0], tensor->op_params[1]);
         } else if (tensor->op == GGML_OP_REPEAT) {
             tensor_clone = ggml_repeat(ggml_ctx, src_clone[0], tensor);
         } else if (tensor->op == GGML_OP_REPEAT_BACK) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/pad_reflect_1d.comp b/ggml/src/ggml-vulkan/vulkan-shaders/pad_reflect_1d.comp
new file mode 100644
index 000000000..2389020fa
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/pad_reflect_1d.comp
@@ -0,0 +1,43 @@
+#version 450
+
+#include "types.glsl"
+#include "generic_unary_head.glsl" // included to use functions like fastdiv etc.
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint p0 = floatBitsToUint(p.param1);
+    const uint p1 = floatBitsToUint(p.param2);
+
+    const uint i3 = fastdiv(idx, p.ne1_012mp, fastdiv_L(p.ne1_Ls, 0));
+    const uint i3_offset = i3 * p.ne12 * p.ne11 * p.ne10;
+
+    const uint i2 = fastdiv(idx - i3_offset, p.ne1_01mp, fastdiv_L(p.ne1_Ls, 1));
+    const uint i2_offset = i2 * p.ne11 * p.ne10;
+
+    const uint i1 = fastdiv(idx - i3_offset - i2_offset, p.ne1_0mp, fastdiv_L(p.ne1_Ls, 2));
+    const uint i0 = idx - i3_offset - i2_offset - i1 * p.ne10;
+
+    uint src_col;
+
+    if (i0 < p0) {
+        src_col = p0 - i0;    // left pad area
+    } else if (i0 < p0 + p.ne00) {
+        src_col = i0 - p0;    // center area
+    } else {
+        src_col = 2u * p.ne00 - 2u - (i0 - p0);  // right pad area
+    }
+
+    const uint src_idx = i3 * p.nb03 + i2 * p.nb02 + i1 * p.nb01 + src_col * p.nb00;
+    const uint d_idx = i3 * p.nb13 + i2 * p.nb12 + i1 * p.nb11 + i0 * p.nb10;
+
+    // copy the computed value to the destination tensor
+    data_d[get_doffset() + d_idx] = D_TYPE(data_a[get_aoffset() + src_idx]);
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index b40a9d291..59c5d72ce 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -922,6 +922,7 @@ void process_shaders() {
     string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
     string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    string_to_spv("pad_reflect_1d_f32", "pad_reflect_1d.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
 
     string_to_spv("concat_i8", "concat.comp", {{"A_TYPE", "uint8_t"}, {"B_TYPE", "uint8_t"}, {"D_TYPE", "uint8_t"}});
     string_to_spv("concat_i16", "concat.comp", {{"A_TYPE", "uint16_t"}, {"B_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index b43058405..fa1484fef 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -4058,6 +4058,41 @@ struct ggml_tensor * ggml_diag_mask_zero_inplace(
     return ggml_diag_mask_zero_impl(ctx, a, n_past, true);
 }
 
+// ggml_clamp
+
+static struct ggml_tensor * ggml_clamp_impl(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 min,
+        float                 max,
+        bool                  inplace) {
+    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    float params[] = { min, max };
+    ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = GGML_OP_CLAMP;
+    result->src[0] = a;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_clamp(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    float                 min,
+    float                 max) {
+    return ggml_clamp_impl(ctx, a, min, max, false);
+}
+
+struct ggml_tensor * ggml_clamp_inplace(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    float                 min,
+    float                 max) {
+    return ggml_clamp_impl(ctx, a, min, max, true);
+}
+
 // ggml_soft_max
 
 static struct ggml_tensor * ggml_soft_max_impl(
@@ -4454,25 +4489,6 @@ struct ggml_tensor * ggml_rope_set_offset(
     return a;
 }
 
-// ggml_clamp
-
-struct ggml_tensor * ggml_clamp(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        float                 min,
-        float                 max) {
-    // TODO: when implement backward, fix this:
-    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
-
-    float params[] = { min, max };
-    ggml_set_op_params(result, params, sizeof(params));
-
-    result->op     = GGML_OP_CLAMP;
-    result->src[0] = a;
-
-    return result;
-}
-
 static int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
     return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
 }
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index 8f6f55519..f236a5d2c 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -3822,7 +3822,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
         MODEL_TENSOR.FFN_EXP_PROBS_B,
-        # NextN/MTP tensors - preserved but unused
+        # NextN/MTP tensors
         MODEL_TENSOR.NEXTN_EH_PROJ,
         MODEL_TENSOR.NEXTN_EMBED_TOKENS,
         MODEL_TENSOR.NEXTN_ENORM,
diff --git a/gpttype_adapter.cpp b/gpttype_adapter.cpp
index 4f66c850a..83d25a1fc 100644
--- a/gpttype_adapter.cpp
+++ b/gpttype_adapter.cpp
@@ -5107,7 +5107,7 @@ std::string gpttype_parse_chat_tool_calls(const std::string & generated_text,
             return "";
         }
 
-        json tools = json::parse(tools_json);
+        common_json tools = common_json::parse(tools_json);
         if(!tools.is_array() || tools.empty())
         {
             return "";
@@ -5135,7 +5135,7 @@ std::string gpttype_parse_chat_tool_calls(const std::string & generated_text,
 
         if(!chat_template_kwargs_json.empty())
         {
-            json kwargs = json::parse(chat_template_kwargs_json);
+            common_json kwargs = common_json::parse(chat_template_kwargs_json);
             if(kwargs.is_object())
             {
                 for(const auto & item : kwargs.items())
@@ -5181,7 +5181,7 @@ std::string gpttype_parse_chat_tool_calls(const std::string & generated_text,
             return "";
         }
 
-        json tool_calls = parsed.to_json_oaicompat().value("tool_calls", json::array());
+        common_json tool_calls = parsed.to_json_oaicompat().value("tool_calls", common_json::array());
         return tool_calls.dump();
     }
     catch(const std::exception & e)
diff --git a/include/llama.h b/include/llama.h
index e071154bb..8890edb22 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -736,7 +736,7 @@ extern "C" {
 
     // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
     // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
-    // seq_id < 0 : match any sequence
+    // seq_id < 0 : match any sequence [TAG_LLAMA_SEQ_ID_NEG]
     // p0 < 0     : [0,  p1]
     // p1 < 0     : [p0, inf)
     LLAMA_API bool llama_memory_seq_rm(
diff --git a/skills/add-new-model/SKILL.md b/skills/add-new-model/SKILL.md
deleted file mode 100644
index f76d1abfd..000000000
--- a/skills/add-new-model/SKILL.md
+++ /dev/null
@@ -1,99 +0,0 @@
----
-name: add-new-model
-description: Guided workflow for adding a new model architecture to llama.cpp. Use when the user wants to add/port a new model architecture.
----
-
-# Add a new model architecture to llama.cpp
-
-This skill walks a contributor through adding a new model architecture. AI-generated code is permitted in this project, so you may write full implementations for the steps below rather than only pointing at patterns - but follow `AGENTS.md`'s AI usage policy throughout:
-
-- The contributor is 100% responsible for every line, however it was produced. They must be able to explain and defend any part of it to a reviewer. Check in with them as you go (don't silently generate everything and hand over a finished diff) so they actually absorb what was written.
-- Before writing code, make sure the contributor owns the design choices for this architecture (which reference model to follow, how non-standard bits like RoPE variants or MoE routing should be handled) - AI accelerates a design the contributor has already made, it doesn't make the design for them.
-- Disclosure is mandatory: any AI-meaningful contribution must be disclosed per the PR template. Remind the contributor of this before they open the PR.
-- Never write the PR description, commit message, GitHub issue/discussion post, or reviewer replies - those must come from the contributor. If asked to commit on their behalf, use `Assisted-by:` (never `Co-authored-by:`) and only after explicit confirmation.
-- If the requested change looks large or introduces a new pattern not covered here, pause and tell the user this kind of change is likely to need prior discussion with maintainers before a PR.
-- Keep the PR self-contained. If the work would require a lot of unconventional changes outside the new model file(s) (e.g. touching shared graph-building code, the sampler, or core APIs in ways other models don't), STOP and tell the contributor to open a discussion/issue first - invasive or excessive changes get closed without full review.
-- Do not bundle unrelated work into this PR - see Step 4 and Step 5 below for the specifics on multimodal and chat-template/parsing work.
-- Never hack around RoPE with a custom sin/cos implementation. Several past PRs tried this and were closed. If the existing `ggml_rope_ext` (see Step 2's RoPE tips) genuinely cannot express what this model needs, the contributor should open an issue to discuss it with maintainers first - not send a PR with a custom RoPE implementation.
-
-Before starting, read `CONTRIBUTING.md`, `AGENTS.md` and `docs/development/HOWTO-add-model.md` if they are not already in context. Also run `git log --oneline -- src/models` and look at at least 3 recent PRs that added a model (their merge commits/diffs) - this shows current convention more reliably than the docs, which can lag behind.
-
-## Step 0 - Scope and dedup check
-
-Ask the contributor:
-1. Which model (HF repo id or name)? Is it text-only or does it have a multimodal (vision/audio) encoder?
-2. Do they already have the HF `config.json`/weights available locally?
-3. Have they checked for an existing PR/issue on this model? Suggest `gh search issues "<model name>"` and `gh search prs "<model name>"` in the `ggml-org/llama.cpp` repo. If an existing PR covers it, the contributor should comment there and collaborate rather than open a duplicate (per CONTRIBUTING.md's AI Usage Policy).
-4. What existing supported architecture is this model closest to (e.g. "Llama-like with sliding window", "MoE like DBRX", "BERT-style encoder")?
-
-If the contributor doesn't know the closest reference architecture, you may grep `conversion/*.py` and `src/models/*.cpp` for architectures with a similar config shape (layer count, head count, MoE expert count, norm placement) and suggest 1-2 candidates - but let the contributor confirm the choice rather than picking one yourself; this choice is a design decision they need to own.
-
-Do not proceed to Step 1 until the contributor has answered these and named a reference architecture.
-
-## Step 1 - Convert the model to GGUF
-
-Follow HOWTO-add-model.md section 1 for the actual touch points (conversion class registration, `constants.py`, `tensor_mapping.py`, etc.) - don't re-derive them here, read them from the doc.
-
-Skill-specific addition: for each touch point, show the contributor the equivalent code in the reference architecture they named in Step 0 before writing the new version, and check that they understand what's different about their model (e.g. non-standard tensor shapes, extra hparams) rather than just copying the pattern silently.
-
-## Step 2 - Define the architecture in llama.cpp
-
-Follow HOWTO-add-model.md section 2 for the actual touch points (`llm_arch` enum, `LLM_ARCH_NAMES`, hparam loading, RoPE type case, etc.), including its "Tips and tricks" section for `ggml_rope_ext` gotchas.
-
-Skill-specific addition: never hack around RoPE with a custom sin/cos implementation - see the RoPE rule above.
-
-## Step 3 - Build the GGML graph
-
-Follow HOWTO-add-model.md section 3 for the actual touch points (`src/models/<name>.cpp` struct, `llama_model_mapping` registration, etc.).
-
-Skill-specific addition: before writing `src/models/<name>.cpp`, read at least 10 other files under `src/models/` (pick a mix, not just the one reference architecture) to confirm the struct layout, naming, and style you're about to write actually matches current convention - the pattern drifts over time and the HOWTO doc can lag behind it.
-
-## Step 4 - Optional: multimodal encoder
-
-Only do this if the contributor flagged a vision/audio encoder in Step 0. Follow HOWTO-add-model.md section 4 and `docs/multimodal.md` for the actual touch points (`MmprojModel` subclass, `clip.cpp`, `mtmd.cpp`, encoder graph in `tools/mtmd/models`, etc.).
-
-Skill-specific addition, and read this carefully: **whether the multimodal encoder can be bundled into the same PR as the base text-model support depends on how conventional the change is.** It's OK to bundle it if the encoder support is conventional - i.e. no new infra or logic is needed, it's just a new cgraph reusing existing preprocessing/projector machinery (e.g. siglip/pixtral/qwen with just a new projector). If it requires anything beyond that - a new preprocessor, non-standard projector logic, or changes to shared `libmtmd` infra/logic - STOP, tell the contributor this is non-conventional, and have them land the text model first with the encoder as a dedicated follow-up PR. Do not let this decision pass silently - call it out explicitly to the contributor before writing any `clip.cpp`/`mtmd.cpp` code.
-
-## Step 5 - Optional: chat template / parsing support
-
-Only do this if the model needs a new built-in chat template (`src/llama-chat.cpp`) or a new output parser (see `docs/development/parsing.md` and `docs/autoparser.md`). If either is needed beyond what a user-supplied Jinja template already covers, treat it as its own dedicated follow-up PR, not part of the base model-support PR - call this out explicitly to the contributor rather than silently bundling it in.
-
-## Common pitfalls (from past PR reviews)
-
-These recur often enough in review comments on past add-model PRs that they're worth checking proactively, not just waiting for a reviewer to catch them:
-
-- Don't validate the same hparam/config assumption in both the Python conversion script and the C++ load path - pick one layer to own the check, duplicating it just adds maintenance surface.
-- Optional hparams that are genuinely absent from some configs (e.g. a shared-expert count) should be read with an explicit optional/fallback accessor, not assumed present.
-- Hparams that are actually load-bearing (the model produces wrong output or crashes without them, e.g. `sliding_window_pattern`, norm-eps) must hard-error if missing, not silently fall back to a default.
-- Don't bake a default chat template into the C++ binary - inject it into the GGUF at conversion time instead, since one `llm_arch` can be reused by multiple fine-tunes with different templates, and a baked-in C++ default fails silently for those.
-- Before writing a dedicated tool-call/output parser, check whether the existing autoparser already handles the template (`llama-debug-template-parser <jinja>` shows what it detects).
-- Marking a custom EOS/closing-tag token as `eot` at conversion time isn't always sufficient - in long/agentic generations a model can emit the closing sequence as literal text instead of the token, so generation never stops on EOG and raw text leaks past the parser. Verify this case, not just the token path.
-- If reusing or aliasing an existing pre-tokenizer for convenience, justify and test that choice explicitly - silent reuse is an easy source of subtle tokenizer bugs.
-- Watch for excessive graph splits caused by building per-layer view/index tensors inside the layer loop - hoist tensors that don't vary per layer out of the loop (relevant if you hit `GGML_SCHED_MAX_SPLIT_INPUTS`).
-- A custom KQ mask fed into flash attention must match FA's expected dtype - cast it to F16 before passing it to `build_attn_mha` when FA is enabled.
-- When padding a custom KV-cache size to an alignment (e.g. `GGML_PAD(..., 256)`), apply the padding after all other size adjustments, not before - otherwise later logic can un-align it again.
-- For non-standard cache/SWA (sliding-window-attention) semantics, override the dedicated hook (e.g. `llama_model_n_swa()`) rather than mutating hparams to fake the behavior - hparams may be read elsewhere for unrelated purposes.
-- Don't ship unfinished or unverified speculative-decoding (e.g. MTP) scaffolding in the base model PR - if it hasn't actually been confirmed to work, pull it out and land it as its own follow-up.
-- Conversion code should call into the base class's existing hparam logic (e.g. `super().set_gguf_parameters()`) rather than re-deriving it - large blocks of code that duplicate what `TextModel`/`MmprojModel` already provide will get flagged as redundant.
-- Do constant tensor modifications (e.g. `norm(1 + weight)`) and permutations/chunking at conversion time, not in the graph - see HOWTO-add-model.md's "Prefer conversion-time tensor modifications" tip (Gemma 3 folds its `1 +` into the weights, Qwen3-Next permutes in `modify_tensors`). Doing these at runtime in the graph is very likely to be rejected as over-complicated; if you genuinely can't do it at conversion time, open a discussion first explaining why rather than implementing it in the graph.
-  - Exception: a plain `weight * scale` with a constant scale is usually better applied at inference time instead of being folded into the weight at conversion. The scale conceptually applies to the activation, not the weight, so folding it in can hurt numerical stability, and it shifts the weight's value range in a way that can make quantization worse.
-
-## Validation checklist
-
-Reference: `examples/model-conversion/README.md`.
-
-1. Convert to GGUF, then inspect/run both the original and converted tensors.
-2. Run logits verification (original vs converted). If this model is a new version of an already-supported family, verify the *previous* version still passes logits verification first - numerical differences may be pre-existing, not caused by the new work. The tools to perform full logits validation are available in `examples/model-conversion`.
-3. Quantize (including QAT variants if relevant) and re-verify.
-4. Run perplexity evaluation (simple and full).
-5. Sanity-check across `tools/cli`, `tools/completion`, `tools/imatrix`, `tools/quantize`, and `tools/server`.
-6. CPU backend first; other backends (CUDA, Metal, ...) can be separate follow-up PRs per `CONTRIBUTING.md`.
-7. Re-review every changed file against the coding/naming guidelines in `AGENTS.md` (and `CONTRIBUTING.md`'s "Coding guidelines"/"Naming guidelines" sections) - this is a separate pass from functional testing and is just as important: no forced line-wrapping, no unicode punctuation, minimal/non-redundant comments, `snake_case` naming (`kebab-case` for file names), matching indentation/brace style, etc.
-
-## Before opening a PR
-
-- Run the `code-review` skill on the diff first - it catches the convention and scope issues reviewers flag most often, and it's recommended to do this locally before pushing the PR.
-- Confirm the contributor can explain every changed line to a reviewer and is prepared to be asked about any of it - this is required regardless of how much of the code was AI-generated.
-- Confirm they did a comprehensive manual review of the full diff, not just a skim.
-- Fill in the AI-disclosure section of `.github/pull_request_template.md` describing how AI was used (do not omit or understate this).
-- Do not write the PR description, commit message, GitHub issue/discussion text, or any reviewer replies yourself - the contributor writes these.
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index 025f9fb54..eecf444fc 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -1060,7 +1060,6 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
         case LLM_ARCH_OLMOE:
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_DEEPSEEK32:
-        case LLM_ARCH_DEEPSEEK4:
         case LLM_ARCH_DOTS3NOTE:
         case LLM_ARCH_GLM_DSA:
         case LLM_ARCH_BITNET:
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index b22623c34..f7dad60e6 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -3228,8 +3228,6 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
 }
 
 size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
-    GGML_UNUSED(seq_id);
-
     if (memory) {
         memory->state_write(io, seq_id, flags);
     }
@@ -3238,8 +3236,6 @@ size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id s
 }
 
 size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
-    GGML_UNUSED(seq_id);
-
     if (memory) {
         memory->state_read(io, seq_id, flags);
     }
diff --git a/src/llama-kv-cache-dsv4.cpp b/src/llama-kv-cache-dsv4.cpp
index 5caa05e8b..948d08146 100644
--- a/src/llama-kv-cache-dsv4.cpp
+++ b/src/llama-kv-cache-dsv4.cpp
@@ -599,6 +599,33 @@ static llama_kv_cache_dsv4_context::comp_plan dsv4_build_comp_plan(
         }
     }
 
+    if (ratio == DSV4_HCA_RATIO && !plan.state_pos.empty() && plan.state_write_idxs.empty()) {
+        assert(kv_size > 0);
+        // the last slot must not be live, or the dummy write would corrupt it;
+        // a full stream implies a completed block, which implies real writes
+        assert(plan.n_kv < (int64_t) kv_size);
+
+        // Keep the compress/write ops in the graph when no HCA block completes
+        // in this ubatch. The dummy block writes to the last cache slot and is
+        // masked out.
+        uint32_t i = 0;
+        while (i < ubatch.n_tokens && ubatch.pos[i] < 0) {
+            ++i;
+        }
+        assert(i < ubatch.n_tokens);
+
+        const llama_seq_id seq_id = ubatch.seq_id[i][0];
+        const int64_t cache_off = dsv4_stream_offset(n_stream, seq_id, kv_size);
+        const int32_t source_idx = state_source_idx(seq_id, ubatch.pos[i]);
+
+        plan.state_write_idxs.push_back(cache_off + kv_size - 1);
+        plan.state_write_pos .push_back(0);
+
+        for (uint32_t j = 0; j < ratio; ++j) {
+            plan.state_read_idxs.push_back(source_idx);
+        }
+    }
+
     if (overlap) {
         // [ all blocks' prev-window indices | all blocks' cur-window indices ]
         plan.state_read_idxs.reserve(overlap_prev_reads.size() + overlap_cur_reads.size());
@@ -608,7 +635,10 @@ static llama_kv_cache_dsv4_context::comp_plan dsv4_build_comp_plan(
                 overlap_cur_reads.begin(), overlap_cur_reads.end());
     }
 
-    plan.n_kv = GGML_PAD(plan.n_kv, 256u);
+    // Keep the mask (and with it the compressed-attention branch) present even
+    // before the first block is visible, so the graph topology never changes.
+    // Padded slots are masked out; comp cache buffers are zero-initialized.
+    plan.n_kv = std::max<int64_t>(GGML_PAD(plan.n_kv, 256u), 256);
 
     std::sort(persist_rows.begin(), persist_rows.end(),
             [](const persist_row & a, const persist_row & b) {
@@ -620,16 +650,26 @@ static llama_kv_cache_dsv4_context::comp_plan dsv4_build_comp_plan(
         plan.state_persist_dst_idxs.push_back(row.dst);
     }
 
-
     if (n_rs_seq > 0) {
-        for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
-            const llama_seq_id seq_id = ubatch.seq_id_unq[s];
-            if (seq_id < 0 || (uint32_t) seq_id >= n_stream) {
-                continue;
+        // Emit restore/snapshot entries for all layout streams so that the
+        // graph tensor sizes do not depend on the ubatch's sequence count.
+        // Streams not present in the ubatch get no-op entries.
+        for (uint32_t stream = 0; stream < n_stream; ++stream) {
+            llama_seq_id seq_id = -1;
+            if (n_stream == 1) {
+                // a unified stream serves any single sequence
+                seq_id = ubatch.n_seqs_unq > 0 ? ubatch.seq_id_unq[0] : -1;
+            } else {
+                for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+                    if (ubatch.seq_id_unq[s] == (llama_seq_id) stream) {
+                        seq_id = ubatch.seq_id_unq[s];
+                        break;
+                    }
+                }
             }
 
-            const int64_t stream_off = dsv4_stream_offset(n_stream, seq_id, state_size);
-            const uint32_t rollback = (uint32_t) seq_id < rs_idx.size() ? rs_idx[seq_id] : 0;
+            const int64_t stream_off = (int64_t) stream*state_size;
+            const uint32_t rollback = seq_id >= 0 && (uint32_t) seq_id < rs_idx.size() ? rs_idx[seq_id] : 0;
             // Keep the restore graph fixed-width when no rollback is pending.
             const int64_t src_plane = rollback > 0 && rollback <= n_rs_seq ? (int64_t) rollback*state_rows : 0;
             for (uint32_t r = 0; r < state_size; ++r) {
@@ -639,35 +679,33 @@ static llama_kv_cache_dsv4_context::comp_plan dsv4_build_comp_plan(
 
             std::vector<uint32_t> token_idxs;
             token_idxs.reserve(ubatch.n_tokens);
-            for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
-                if (dsv4_token_has_seq(ubatch, i, seq_id)) {
-                    token_idxs.push_back(i);
+            if (seq_id >= 0) {
+                for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
+                    if (dsv4_token_has_seq(ubatch, i, seq_id)) {
+                        token_idxs.push_back(i);
+                    }
                 }
             }
-            if (token_idxs.empty()) {
-                continue;
-            }
 
             const uint32_t n_seq_tokens = (uint32_t) token_idxs.size();
             const int64_t scratch_off = (int64_t) state_rows*(1 + n_rs_seq);
             for (uint32_t d = 1; d <= n_rs_seq; ++d) {
                 const int64_t dst_plane = (int64_t) d*state_rows;
+                const uint32_t prefix = d <= n_seq_tokens ? n_seq_tokens - d : 0;
 
                 for (uint32_t r = 0; r < state_size; ++r) {
-                    int32_t src;
-                    if (d <= n_seq_tokens) {
-                        const uint32_t prefix = n_seq_tokens - d;
-                        src = (int32_t) (stream_off + r);
+                    int32_t src = (int32_t) (stream_off + r);
 
-                        for (uint32_t j = 0; j < prefix; ++j) {
-                            const uint32_t i_tok = token_idxs[j];
-                            if (ubatch.pos[i_tok] >= 0 && (uint32_t) (ubatch.pos[i_tok]%state_size) == r) {
-                                src = (int32_t) (scratch_off + i_tok);
-                            }
+                    for (uint32_t j = 0; j < prefix; ++j) {
+                        const uint32_t i_tok = token_idxs[j];
+                        if (ubatch.pos[i_tok] >= 0 && (uint32_t) (ubatch.pos[i_tok]%state_size) == r) {
+                            src = (int32_t) (scratch_off + i_tok);
                         }
-                    } else {
-                        const int64_t src_plane = (int64_t) (d - n_seq_tokens)*state_rows;
-                        src = (int32_t) (src_plane + stream_off + r);
+                    }
+
+                    if (n_seq_tokens == 0) {
+                        // no-op: copy the snapshot plane onto itself
+                        src = (int32_t) (dst_plane + stream_off + r);
                     }
 
                     plan.state_snapshot_src_idxs.push_back(src);
@@ -683,10 +721,16 @@ static llama_kv_cache_dsv4_context::comp_plan dsv4_build_comp_plan(
     }();
 
     if (debug) {
-        LLAMA_LOG_INFO("%s: ratio=%u, n_tokens=%u, state_persist_dst=%s, state_write_pos=%s\n",
-                __func__, ratio, ubatch.n_tokens,
+        LLAMA_LOG_DEBUG("%s: ratio=%u, n_tokens=%u, n_seqs_unq=%u, state_persist_dst=%s, state_write_pos=%s\n",
+                __func__, ratio, ubatch.n_tokens, ubatch.n_seqs_unq,
                 dsv4_plan_positions(plan.state_persist_dst_idxs).c_str(),
                 dsv4_plan_positions(plan.state_write_pos).c_str());
+        for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+            const llama_seq_id seq_id = ubatch.seq_id_unq[s];
+            const uint32_t rollback = seq_id >= 0 && (uint32_t) seq_id < rs_idx.size() ? rs_idx[seq_id] : 0;
+            LLAMA_LOG_DEBUG("%s:   seq %d pos [%d, %d] rollback=%u\n", __func__, seq_id,
+                    ubatch.pos[0], ubatch.pos[ubatch.n_tokens - 1], rollback);
+        }
     }
 
     return plan;
@@ -704,8 +748,17 @@ static std::vector<llama_kv_cache_dsv4_context::comp_plan> dsv4_build_comp_plans
     std::vector<llama_kv_cache_dsv4_context::comp_plan> plans;
     plans.reserve(ubatches.size());
 
+    // the first ubatch touching a seq consumes its rollback restore
+    std::vector<uint32_t> rs(rs_idx);
     for (const llama_ubatch & ubatch : ubatches) {
-        plans.push_back(dsv4_build_comp_plan(ubatch, ratio, overlap, state_size, kv_size, n_stream, n_rs_seq, rs_idx));
+        plans.push_back(dsv4_build_comp_plan(ubatch, ratio, overlap, state_size, kv_size, n_stream, n_rs_seq, rs));
+
+        for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+            const llama_seq_id seq_id = ubatch.seq_id_unq[s];
+            if (seq_id >= 0 && (size_t) seq_id < rs.size()) {
+                rs[seq_id] = 0;
+            }
+        }
     }
 
     return plans;
@@ -803,16 +856,15 @@ static llama_kv_cache_dsv4_context::comp_plan dsv4_build_reserve_comp_plan(
         return plan;
     }
 
-    const uint32_t n_seqs       = std::max<uint32_t>(1, ubatch.n_seqs);
-    const uint32_t n_seq_tokens = std::max<uint32_t>(1, ubatch.n_seq_tokens);
-    const uint64_t n_blocks_u64 = (uint64_t) n_seqs*((n_seq_tokens + ratio - 1)/ratio);
-    const size_t n_blocks = (size_t) std::max<uint64_t>(1, n_blocks_u64);
-    GGML_ASSERT((uint64_t) n_blocks == std::max<uint64_t>(1, n_blocks_u64));
+    // worst case over every seq split: sum of per-seq ceil(tokens/ratio) is at
+    // most floor(n_tokens/ratio) + n_seqs
+    const uint32_t n_seqs = std::max<uint32_t>(1, ubatch.n_seqs);
+    const size_t n_blocks = (size_t) ubatch.n_tokens/ratio + n_seqs;
 
     const uint64_t state_rows = (uint64_t) state_size*n_stream;
     const size_t n_persist = (size_t) std::min<uint64_t>(ubatch.n_tokens, state_rows);
-    const size_t n_restore = n_rs_seq > 0 ? (size_t) state_size*std::max<uint32_t>(1, ubatch.n_seqs_unq) : 0;
-    const size_t n_snapshot = (size_t) n_rs_seq*state_size*std::max<uint32_t>(1, ubatch.n_seqs_unq);
+    const size_t n_restore = n_rs_seq > 0 ? (size_t) state_size*n_stream : 0;
+    const size_t n_snapshot = (size_t) n_rs_seq*state_size*n_stream;
 
     plan.state_pos .resize(ubatch.n_tokens);
     plan.state_persist_src_idxs.resize(n_persist);
@@ -1356,7 +1408,9 @@ llama_memory_context_ptr llama_kv_cache_dsv4::init_batch(
             if (has_coupled) {
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                ubatch = balloc.split_equal(n_ubatch, raw_per_seq || comp_per_seq, 0);
+                // [TAG_RECURRENT_ROLLBACK_SPLITS]
+                // the trailing (1 + n_rs_seq) tokens of each seq must stay in the same ubatch
+                ubatch = balloc.split_equal(n_ubatch, raw_per_seq || comp_per_seq, n_rs_seq > 0 ? n_rs_seq + 1 : 0);
             }
 
             if (ubatch.n_tokens == 0) {
@@ -1433,6 +1487,11 @@ bool llama_kv_cache_dsv4::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1
             return false;
         }
 
+        // pending rollback is single-use: stacked partial removals don't compose
+        if (rs_idx[seq_id] != 0) {
+            return false;
+        }
+
         const bool res = kv_raw->seq_rm(seq_id, p0, p1);
         if (res) {
             rs_idx[seq_id] = (uint32_t) rollback;
@@ -1594,9 +1653,7 @@ void llama_kv_cache_dsv4::state_read(llama_io_read_i & io, llama_seq_id seq_id,
     kv_raw->state_read(io, seq_id, flags);
 
     if (!partial_only) {
-        kv_csa->clear(true);
-        kv_hca->clear(true);
-        kv_lid->clear(true);
+        clear_compressed(seq_id, true);
 
         dsv4_state_read_k_cache(io, kv_csa.get(), seq_id, flags);
         dsv4_state_read_k_cache(io, kv_hca.get(), seq_id, flags);
@@ -1680,6 +1737,7 @@ void llama_kv_cache_dsv4::clear_compressed(llama_seq_id seq_id, bool data) {
             kv->seq_rm(seq_id, -1, -1);
 
             if (data) {
+                //TODO: do not clear the kv-cache during `seq_rm`, ref: https://github.com/ggml-org/llama.cpp/pull/26490#discussion_r3798143663
                 for (uint32_t il : kv->get_layer_ids()) {
                     dsv4_clear_tensor_stream(kv->get_k_storage(il), (uint32_t) seq_id);
                 }
diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
index 01ef0255c..ea1d720cb 100644
--- a/src/llama-kv-cache.cpp
+++ b/src/llama-kv-cache.cpp
@@ -383,6 +383,7 @@ bool llama_kv_cache::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
         return true;
     }
 
+    // TODO: fix incosistent handling of `seq_id < 0` and `seq_id == -1` in the codebase [TAG_LLAMA_SEQ_ID_NEG]
     GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));
 
     if (p0 < 0) {
@@ -2048,6 +2049,7 @@ void llama_kv_cache::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama
 
     GGML_UNUSED(flags);
 
+    // TODO: fix incosistent handling of `seq_id < 0` and `seq_id == -1` in the codebase [TAG_LLAMA_SEQ_ID_NEG]
     GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));
 
     uint32_t n_stream_cur;
diff --git a/src/llama-memory-recurrent.cpp b/src/llama-memory-recurrent.cpp
index 8e8d654b5..c4e64d7d3 100644
--- a/src/llama-memory-recurrent.cpp
+++ b/src/llama-memory-recurrent.cpp
@@ -158,13 +158,14 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         p1 = std::numeric_limits<llama_pos>::max();
     }
 
+    if ((uint32_t) seq_id >= this->n_seq_max) {
+        LLAMA_LOG_ERROR("%s: invalid seq_id (%d) - larger than n_seq_max (%d)\n", __func__, seq_id, this->n_seq_max);
+        return false;
+    }
+
     const bool rm_all = p0 == 0 && p1 == std::numeric_limits<llama_pos>::max();
     if (rm_all) {
-        if (seq_id >= 0) {
-            set_rs_idx(seq_id, 0);
-        } else {
-            std::fill(rs_idx.begin(), rs_idx.end(), 0);
-        }
+        set_rs_idx(seq_id, 0);
     }
 
     // models like Mamba or RWKV can't have a state partially erased at the end
@@ -181,7 +182,9 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
             // partial rollback via per-token snapshot index (bounded by n_rs_seq)
             if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) {
                 const llama_pos rollback = cell.pos - (p0 - 1);
-                if (rollback >= 1 && rollback <= (llama_pos) n_rs_seq) {
+                // pending rollback is single-use
+                const bool pending = rs_idx[seq_id] != 0;
+                if (!pending && rollback >= 1 && rollback <= (llama_pos) n_rs_seq) {
                     set_rs_idx(seq_id, (uint32_t) rollback);
                     cell.pos = p0 - 1;
                     return true;
@@ -390,10 +393,17 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const {
 }
 
 void llama_memory_recurrent::set_rs_idx(llama_seq_id seq_id, uint32_t idx) {
-    if (seq_id < 0 || (size_t) seq_id >= rs_idx.size()) {
+    if (seq_id < 0) {
+        std::fill(rs_idx.begin(), rs_idx.end(), 0);
         return;
     }
-    rs_idx[seq_id] = (idx > n_rs_seq) ? n_rs_seq : idx;
+
+    assert(n_seq_max == rs_idx.size());
+
+    GGML_ASSERT((uint32_t) seq_id < n_seq_max);
+    GGML_ASSERT(idx <= n_rs_seq);
+
+    rs_idx[seq_id] = idx;
 }
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_memory_recurrent::memory_breakdown() const {
@@ -742,6 +752,7 @@ void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq
     uint32_t cell_range_begin = size;
     for (uint32_t i = 0; i < size; ++i) {
         const auto & cell = cells[i];
+        // TODO: fix incosistent handling of `seq_id < 0` and `seq_id == -1` in the codebase [TAG_LLAMA_SEQ_ID_NEG]
         if ((seq_id == -1 && !cell.is_empty()) || cell.has_seq_id(seq_id)) {
             ++cell_count;
             uint32_t rs_idx_cur = 0;
@@ -827,6 +838,7 @@ void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_i
     }
 
     if (!res) {
+        // TODO: fix incosistent handling of `seq_id < 0` and `seq_id == -1` in the codebase [TAG_LLAMA_SEQ_ID_NEG]
         if (seq_id == -1) {
             clear(true);
         } else {
@@ -836,11 +848,7 @@ void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_i
     }
 
     if (n_rs_seq != 0) {
-        if (seq_id == -1) {
-            std::fill(rs_idx.begin(), rs_idx.end(), 0);
-        } else {
-            set_rs_idx(seq_id, 0);
-        }
+        set_rs_idx(seq_id, 0);
     }
 }
 
diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cpp
index 0d39e6de8..9adaa93f6 100644
--- a/src/llama-model-saver.cpp
+++ b/src/llama-model-saver.cpp
@@ -293,6 +293,21 @@ void llama_model_saver::add_kv_from_model() {
     add_kv(LLM_KV_ATTENTION_INDEXER_LOCAL_BLOCKS,    hparams.indexer_local_blocks);
     add_kv(LLM_KV_ATTENTION_INDEXER_TYPES,           hparams.is_indexer_full_impl, true);
     add_kv(LLM_KV_ATTENTION_RECURRENT_LAYERS,        hparams.is_recr_impl, true);
+    add_kv(LLM_KV_ATTENTION_OUTPUT_GROUP_COUNT,      hparams.dsv4_o_group_count);
+    add_kv(LLM_KV_ATTENTION_OUTPUT_LORA_RANK,        hparams.dsv4_o_lora_rank);
+    add_kv(LLM_KV_ATTENTION_COMPRESS_ROPE_FREQ_BASE, hparams.dsv4_compress_rope_base);
+    if (model->arch == LLM_ARCH_DEEPSEEK4 || hparams.dsv4_hc_mult > 0) {
+        // the loader requires one compress ratio per layer, including nextn layers
+        const std::vector<uint32_t> compress_ratios(
+                hparams.dsv4_compress_ratios.begin(), hparams.dsv4_compress_ratios.begin() + hparams.n_layer_all);
+        add_kv(LLM_KV_ATTENTION_COMPRESS_RATIOS,         compress_ratios);
+    } else {
+        add_kv(LLM_KV_ATTENTION_COMPRESS_RATIOS,         hparams.dsv4_compress_ratios, true);
+    }
+    add_kv(LLM_KV_HYPER_CONNECTION_COUNT,               hparams.dsv4_hc_mult);
+    add_kv(LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS, hparams.dsv4_hc_sinkhorn_iters);
+    add_kv(LLM_KV_HYPER_CONNECTION_EPSILON,             hparams.dsv4_hc_eps);
+    add_kv(LLM_KV_HASH_LAYER_COUNT,                     hparams.dsv4_hash_layer_count);
 
     const float rope_scaling_factor = hparams.rope_freq_scale_train == 1.0f ? 0.0f : 1.0f/hparams.rope_freq_scale_train;
 
@@ -417,11 +432,16 @@ void llama_model_saver::add_tensors_from_model() {
     add_tensor(model->output_s);
     add_tensor(model->output_in_s);
     add_tensor(model->output_res_score);
+    add_tensor(model->nextn_proj_pre);
+    add_tensor(model->nextn_proj_post);
     add_tensor(model->cls);
     add_tensor(model->cls_b);
     add_tensor(model->cls_out);
     add_tensor(model->cls_out_b);
     add_tensor(model->cls_norm);
+    add_tensor(model->hc_head_fn);
+    add_tensor(model->hc_head_base);
+    add_tensor(model->hc_head_scale);
 
     for (const struct llama_layer & layer : model->layers) {
         for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 5b0e71847..9312eb747 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -516,6 +516,8 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
     const llama_meta_device_get_split_state_userdata * ud = (const llama_meta_device_get_split_state_userdata *) userdata;
     const llama_hparams & hparams = ud->model->hparams;
     const std::string tensor_name = tensor->name;
+    const bool is_dsv4 = ud->model->arch == LLM_ARCH_DEEPSEEK4 ||
+        (ud->model->arch == LLM_ARCH_DFLASH && hparams.dsv4_hc_mult > 0);
 
     static const std::regex pattern_q_weight        ("blk\\.\\d*\\.attn_q.weight");
     static const std::regex pattern_kv_weight       ("blk\\.\\d*\\.attn_(k|v).weight");
@@ -525,9 +527,13 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
     static const std::regex pattern_qkv_bias        ("blk\\.\\d*\\.attn_qkv.bias");
     static const std::regex pattern_qk_norm         ("blk\\.\\d*\\.attn_(q|k)_norm\\.weight");
     static const std::regex pattern_kv_cache        ("cache_(k|v)_l\\d*");
+    static const std::regex pattern_dsv4_state      ("dsv4_(csa|hca|lid)_state_(kv|score)_l\\d*");
     static const std::regex pattern_attn_sinks      ("blk\\.\\d*\\.attn_sinks.weight");
     static const std::regex pattern_attn_out_weight ("blk\\.\\d*\\.attn_output.weight");
     static const std::regex pattern_attn_out_bias   ("blk\\.\\d*\\.attn_output.bias");
+    static const std::regex pattern_attn_out_a_weight("blk\\.\\d*\\.attn_output_a\\.weight");
+    static const std::regex pattern_attn_out_b_weight("blk\\.\\d*\\.attn_output_b\\.weight");
+    static const std::regex pattern_attn_q_b_weight ("blk\\.\\d*\\.attn_q_b\\.weight");
     static const std::regex pattern_attn_gate_weight("blk\\.\\d*\\.attn_gate.weight");
 
     static const std::regex pattern_ssm_dt          ("blk\\.\\d*\\.ssm_dt.bias");
@@ -546,8 +552,11 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
     static const std::regex pattern_ffn_gate_bias     ("blk\\.\\d*\\.ffn_gate(_exps)?.bias");
     static const std::regex pattern_ffn_gate_up_weight("blk\\.\\d*\\.ffn_gate_up(_exps)?.weight");
     static const std::regex pattern_ffn_down_weight   ("blk\\.\\d*\\.ffn_down(_exps)?.weight");
-    static const std::regex pattern_ffn_down_bias     ("blk\\.\\d*\\.ffn_down.bias");
-    static const std::regex pattern_ffn_down_exps_bias("blk\\.\\d*\\.ffn_down_exps.bias");
+    static const std::regex pattern_ffn_down_bias         ("blk\\.\\d*\\.ffn_down.bias");
+    static const std::regex pattern_ffn_down_exps_bias    ("blk\\.\\d*\\.ffn_down_exps.bias");
+    static const std::regex pattern_ffn_up_shexp_weight   ("blk\\.\\d*\\.ffn_up_shexp.weight");
+    static const std::regex pattern_ffn_gate_shexp_weight ("blk\\.\\d*\\.ffn_gate_shexp.weight");
+    static const std::regex pattern_ffn_down_shexp_weight ("blk\\.\\d*\\.ffn_down_shexp.weight");
 
     static const std::regex pattern_output_weight("output\\.weight");
     static const std::regex pattern_output_bias  ("output\\.bias");
@@ -604,6 +613,32 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
     };
 
     auto get_tensor_config = [&]() -> tensor_config {
+        if (is_dsv4) {
+            if (std::regex_match(tensor_name, pattern_kv_cache) ||
+                    std::regex_match(tensor_name, pattern_dsv4_state)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+            }
+            if (std::regex_match(tensor_name, pattern_attn_sinks)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_0, "attn_output_a.weight");
+            }
+            if (std::regex_match(tensor_name, pattern_attn_q_b_weight)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1, "attn_output_a.weight");
+            }
+            if (std::regex_match(tensor_name, pattern_attn_out_a_weight)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_2);
+            }
+            if (std::regex_match(tensor_name, pattern_attn_out_b_weight)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_0);
+            }
+            if (std::regex_match(tensor_name, pattern_ffn_up_shexp_weight) ||
+                    std::regex_match(tensor_name, pattern_ffn_gate_shexp_weight)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1, "ffn_down_shexp.weight");
+            }
+            if (std::regex_match(tensor_name, pattern_ffn_down_shexp_weight)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_0, "ffn_down_shexp.weight");
+            }
+        }
+
         // standard attention
         if (std::regex_match(tensor_name, pattern_q_weight) || std::regex_match(tensor_name, pattern_kv_weight)) {
             return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1, "attn_output.weight", "ssm_out.weight");
@@ -671,11 +706,14 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
             return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_MIRRORED);
         }
         if (std::regex_match(tensor_name, pattern_ffn_down_exps_bias)) {
-            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_PARTIAL);
+            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_PARTIAL, "ffn_down_exps.weight");
         }
 
         // output
         if (std::regex_match(tensor_name, pattern_output_weight)) {
+            if (is_dsv4) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+            }
             return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1);
         }
         if (std::regex_match(tensor_name, pattern_output_bias)) {
@@ -705,6 +743,9 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
                     GGML_ASSERT(tensor->ne[axis] == 2*key_dim + value_dim);
                     return {{key_dim, 2}, {value_dim, 1}};
                 }
+                if (std::regex_match(tensor_name, pattern_r_cache)) {
+                    return {{key_dim * (hparams.ssm_d_conv - 1), 2}, {value_dim * (hparams.ssm_d_conv - 1), 1}};
+                }
             } else {
                 const int64_t head_ratio = n_v_heads / n_k_heads;
                 if (std::regex_match(tensor_name, pattern_qkv_weight) || std::regex_match(tensor_name, pattern_ssm_conv1d)) {
@@ -793,12 +834,34 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
                 blck_size_perf *= 2;
             }
 
+            const int64_t granularity_q    = std::lcm(n_embd_q, blck_size_perf);
+            const int64_t granularity_head = granularity_q / hparams.n_embd_head_k(il); // for tensors with one value per head
             if (std::regex_match(tensor_name, pattern_attn_sinks)) {
                 GGML_ASSERT(segments.size() == 1);
-                return {std::lcm(n_embd_q, blck_size_perf)/n_embd_q * n_gqa};
+                if (is_dsv4) {
+                    return {hparams.n_head(il) / hparams.dsv4_o_group_count};
+                }
+                return {granularity_head};
             }
 
-            const int64_t granularity_q = std::lcm(n_embd_q, blck_size_perf);
+            if (is_dsv4) {
+                if (std::regex_match(tensor_name, pattern_attn_q_b_weight)) {
+                    GGML_ASSERT(segments.size() == 1);
+                    // the grouped output projection requires each device to hold whole groups of heads
+                    const int64_t n_head_group = hparams.n_head(il) / hparams.dsv4_o_group_count;
+                    return {n_head_group * hparams.n_embd_head_k(il)};
+                }
+                if (std::regex_match(tensor_name, pattern_attn_out_a_weight)) {
+                    GGML_ASSERT(segments.size() == 1);
+                    return {1};
+                }
+                if (std::regex_match(tensor_name, pattern_attn_out_b_weight)) {
+                    GGML_ASSERT(segments.size() == 1);
+                    // the boundaries must align with wo_a's per-group split, so quant blocks must not straddle groups
+                    GGML_ASSERT(hparams.dsv4_o_lora_rank % blck_size == 0);
+                    return {hparams.dsv4_o_lora_rank};
+                }
+            }
             if (std::regex_match(tensor_name, pattern_q_weight) || std::regex_match(tensor_name, pattern_q_bias)) {
                 GGML_ASSERT(segments.size() == 1);
                 // some models have Q gate tensors, for those cases the granularity needs to be doubled:
@@ -811,6 +874,13 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
                 GGML_ASSERT(segments.size() == 1);
                 return {granularity_q};
             }
+            if (std::regex_match(tensor_name, pattern_attn_gate_weight)) {
+                GGML_ASSERT(segments.size() == 1);
+                if (tensor->ne[1] == hparams.n_head(il)) {
+                    return {granularity_head};
+                }
+                return {granularity_q};
+            }
 
             const int64_t granularity_kv = granularity_q / n_gqa;
             if (std::regex_match(tensor_name, pattern_kv_weight) ||
@@ -828,7 +898,11 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
         // FFN
         if (std::regex_match(tensor_name, pattern_ffn_up_weight) || std::regex_match(tensor_name, pattern_ffn_up_bias) ||
                 std::regex_match(tensor_name, pattern_ffn_gate_weight) || std::regex_match(tensor_name, pattern_ffn_gate_bias) ||
-                std::regex_match(tensor_name, pattern_ffn_gate_up_weight) || std::regex_match(tensor_name, pattern_ffn_down_weight)) {
+                std::regex_match(tensor_name, pattern_ffn_gate_up_weight) ||
+                std::regex_match(tensor_name, pattern_ffn_down_weight) ||
+                std::regex_match(tensor_name, pattern_ffn_up_shexp_weight) ||
+                std::regex_match(tensor_name, pattern_ffn_gate_shexp_weight) ||
+                std::regex_match(tensor_name, pattern_ffn_down_shexp_weight)) {
             const int64_t blck_size_perf = std::lcm(blck_size, 128);
             GGML_ASSERT(segments.size() == 1);
             return {blck_size_perf};
@@ -879,6 +953,16 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
         memset(split_state.ne, 0, sizeof(split_state.ne));
         split_state.nr[0] = 1;
         split_state.n_segments = 1;
+        if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
+            GGML_ASSERT(tc.tensor_axis_0 != tensor);
+            const ggml_backend_meta_split_state source_split_state = llama_meta_device_get_split_state(tc.tensor_axis_0, userdata);
+            GGML_ASSERT(source_split_state.axis >= 0 && source_split_state.axis < GGML_MAX_DIMS);
+            for (size_t j = 0; j < ud->n_devices; j++) {
+                for (size_t is = 0; is < source_split_state.n_segments; is++) {
+                    split_state.ne[j] += source_split_state.ne[is*ud->n_devices + j] * source_split_state.nr[is];
+                }
+            }
+        }
     }
     return split_state;
     GGML_UNUSED(userdata);
diff --git a/src/models/deepseek4.cpp b/src/models/deepseek4.cpp
index 1c278e435..fc816e2ae 100644
--- a/src/models/deepseek4.cpp
+++ b/src/models/deepseek4.cpp
@@ -1,3 +1,4 @@
+#include "llama-hparams.h"
 #include "models.h"
 
 #include "llama-kv-cache-dsv4.h"
@@ -58,6 +59,7 @@ void llama_model_deepseek4::load_arch_hparams(llama_model_loader & ml) {
     if (n_compress_ratios < hparams.n_layer_all) {
         throw std::runtime_error("DeepSeek-V4 compress_ratios is shorter than block_count");
     }
+    GGML_ASSERT(n_compress_ratios <= LLAMA_MAX_LAYERS);
     ml.get_arr(LLM_KV_ATTENTION_COMPRESS_RATIOS, hparams.dsv4_compress_ratios);
 
     ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func);
diff --git a/src/models/dflash.cpp b/src/models/dflash.cpp
index d3c919b35..ff40c16b2 100644
--- a/src/models/dflash.cpp
+++ b/src/models/dflash.cpp
@@ -117,6 +117,10 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
     output_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), { n_embd }, 0); // encoder hidden_norm (after fc)
     output_norm     = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM,    "weight"), { n_embd }, 0); // decoder final norm
 
+    // optional: reduced-vocab drafts ship their own lm head, full-vocab drafts can share the target's via ctx_other
+    // a draft with its own embeddings + head references no target tensors and can run on devices the target does not use (e.g. -devd with a tensor-split target)
+    output   = create_tensor(tn(LLM_TENSOR_OUTPUT,     "weight"), { n_embd, n_vocab_draft }, TENSOR_NOT_REQUIRED);
+
     if (hparams.dsv4_hc_mult > 0) {
         const int64_t q_lora_rank     = hparams.n_lora_q;
         const int64_t n_ff_exp        = hparams.n_ff_exp;
@@ -167,9 +171,6 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
         return;
     }
 
-    // optional: reduced-vocab drafts ship their own, full-vocab drafts share the target's via ctx_other
-    output   = create_tensor(tn(LLM_TENSOR_OUTPUT,     "weight"), { n_embd, n_vocab_draft }, TENSOR_NOT_REQUIRED);
-
     for (int i = 0; i < n_layer; ++i) {
         auto & layer = layers[i];
 
diff --git a/src/models/glm4-moe.cpp b/src/models/glm4-moe.cpp
index 8cde66978..83ea7f8ac 100644
--- a/src/models/glm4-moe.cpp
+++ b/src/models/glm4-moe.cpp
@@ -29,10 +29,19 @@ void llama_model_glm4_moe::load_arch_hparams(llama_model_loader & ml) {
     }
 }
 
-void llama_model_glm4_moe::load_arch_tensors(llama_model_loader &) {
+void llama_model_glm4_moe::load_arch_tensors(llama_model_loader & ml) {
     LLAMA_LOAD_LOCALS;
     const int64_t n_expert_shared = hparams.n_expert_shared;
 
+    const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);
+    const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";
+    const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);
+    const int trunk_flags = mtp_only  ? TENSOR_NOT_REQUIRED : 0;
+    int       mtp_flags   = trunk_only ? TENSOR_NOT_REQUIRED : 0;
+
+    if (!ml.load_mtp) {
+        mtp_flags |= TENSOR_SKIP;
+    }
 
     GGML_ASSERT(hparams.n_expert > 0 && "n_expert must be > 0 for GLM4_MOE MoE layers");
     GGML_ASSERT(hparams.n_expert_used > 0 && "n_expert_used must be > 0 for GLM4_MOE MoE layers");
@@ -47,16 +56,9 @@ void llama_model_glm4_moe::load_arch_tensors(llama_model_loader &) {
         output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
     }
 
-    // Load ALL tensors including NextN layer to satisfy total tensor count
-    // but only PROCESS up to last layer (skipping final NextN layer) in forward pass
     for (int i = 0; i < n_layer_all; ++i) {
-        int flags = 0;
-        if (i >= n_layer) {
-            // skip all tensors in the NextN layers
-            flags |= TENSOR_SKIP;
-        }
-
         auto & layer = layers[i];
+        const int flags = i < n_layer ? trunk_flags : mtp_flags;
 
         layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, flags);
 
@@ -110,24 +112,186 @@ void llama_model_glm4_moe::load_arch_tensors(llama_model_loader &) {
             layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), { n_embd, n_ff }, flags);
         }
 
-        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+        // NextN/MTP tensors
         if (i >= n_layer) {
             layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
             layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
             layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
 
             // Optional tensors
-            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
-            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
-            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);
+            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
+            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
+            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, TENSOR_NOT_REQUIRED | flags);
         }
     }
 }
 
 std::unique_ptr<llm_graph_context> llama_model_glm4_moe::build_arch_graph(const llm_graph_params & params) const {
+    if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {
+        return std::make_unique<graph_mtp>(*this, params);
+    }
     return std::make_unique<graph>(*this, params);
 }
 
+llama_model_glm4_moe::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params)
+    : llm_graph_context(params) {
+    GGML_ASSERT(hparams.n_layer_nextn > 0 && "GLM4_MOE MTP requires n_layer_nextn > 0");
+    GGML_ASSERT(hparams.n_layer_nextn == 1 && "GLM4_MOE MTP currently only supports a single MTP block");
+
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+
+    const int il = hparams.n_layer() + cparams.nextn_layer_offset;
+    GGML_ASSERT(cparams.nextn_layer_offset >= 0 &&
+                cparams.nextn_layer_offset < (int) hparams.n_layer_nextn &&
+                "nextn_layer_offset out of range [0, n_layer_nextn)");
+
+    const auto & layer = model.layers[il];
+
+    GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj");
+    GGML_ASSERT(layer.nextn.enorm   && "MTP block missing nextn.enorm");
+    GGML_ASSERT(layer.nextn.hnorm   && "MTP block missing nextn.hnorm");
+    GGML_ASSERT(layer.ffn_gate_inp  && "MTP block missing ffn_gate_inp");
+
+    auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd);
+
+    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    ggml_set_input(inp->tokens);
+
+    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);
+    ggml_set_input(inp->embd);
+
+    ggml_tensor * tok_embd;
+    if (ubatch.token) {
+        ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;
+        tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);
+    } else {
+        tok_embd = inp->embd;
+    }
+    cb(tok_embd, "mtp_tok_embd", il);
+
+    inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
+    ggml_set_input(inp->h);
+    ggml_set_name(inp->h, "mtp_h_input");
+
+    ggml_tensor * h_embd = inp->h;
+
+    res->add_input(std::move(inp));
+
+    ggml_tensor * inp_pos     = build_inp_pos();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
+    cb(h_norm, "mtp_hnorm", il);
+
+    ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);
+    cb(e_norm, "mtp_enorm", il);
+
+    ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, 0);
+    cb(concat, "mtp_concat", il);
+
+    ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s);
+    cb(cur, "mtp_eh_proj", il);
+
+    ggml_tensor * inpSA = cur;
+
+    cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);
+    cb(cur, "mtp_attn_norm", il);
+
+    auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur,
+            n_embd_head, n_head, n_head_kv, il);
+
+    if (layer.attn_q_norm) {
+        Qcur = build_norm(Qcur, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il);
+        cb(Qcur, "mtp_Qcur_normed", il);
+    }
+    if (layer.attn_k_norm) {
+        Kcur = build_norm(Kcur, layer.attn_k_norm, nullptr, LLM_NORM_RMS, il);
+        cb(Kcur, "mtp_Kcur_normed", il);
+    }
+
+    Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,
+            rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+
+    Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,
+            rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+
+    cb(Qcur, "mtp_Qcur", il);
+    cb(Kcur, "mtp_Kcur", il);
+    cb(Vcur, "mtp_Vcur", il);
+
+    cur = build_attn(inp_attn,
+            layer.wo, nullptr, layer.wo_s,
+            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr,
+            1.0f / sqrtf(float(n_embd_head)), il);
+    cb(cur, "mtp_attn_out", il);
+
+    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+    cb(ffn_inp, "mtp_ffn_inp", il);
+
+    cur = build_norm(ffn_inp, layer.attn_post_norm, nullptr, LLM_NORM_RMS, il);
+    cb(cur, "mtp_post_attn_norm", il);
+
+    ggml_tensor * routed_out = build_moe_ffn(cur,
+            layer.ffn_gate_inp,
+            layer.ffn_up_exps,
+            layer.ffn_gate_exps,
+            layer.ffn_down_exps,
+            layer.ffn_exp_probs_b,
+            n_expert, n_expert_used,
+            LLM_FFN_SILU, hparams.expert_weights_norm,
+            hparams.expert_weights_scale,
+            (llama_expert_gating_func_type) hparams.expert_gating_func,
+            il);
+    cb(routed_out, "mtp_ffn_moe_out", il);
+
+    ggml_tensor * shared_out = build_ffn(cur,
+            layer.ffn_up_shexp,   nullptr, nullptr,
+            layer.ffn_gate_shexp, nullptr, nullptr,
+            layer.ffn_down_shexp, nullptr, nullptr,
+            nullptr,
+            LLM_FFN_SILU, LLM_FFN_PAR, il);
+    cb(shared_out, "mtp_ffn_shexp_out", il);
+
+    cur = ggml_add(ctx0, routed_out, shared_out);
+    cb(cur, "mtp_ffn_out", il);
+
+    cur = ggml_add(ctx0, cur, ffn_inp);
+    cb(cur, "mtp_post_ffn", il);
+
+    ggml_tensor * head_norm_w = layer.nextn.shared_head_norm
+            ? layer.nextn.shared_head_norm
+            : model.output_norm;
+    GGML_ASSERT(head_norm_w && "GLM4_MOE MTP: missing both nextn.shared_head_norm and output_norm");
+
+    cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    if (inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+    cb(cur, "mtp_shared_head_norm", -1);
+
+    ggml_tensor * head_w = layer.nextn.shared_head_head
+            ? layer.nextn.shared_head_head
+            : model.output;
+    ggml_tensor * head_s = layer.nextn.shared_head_head
+            ? layer.nextn.shared_head_head_s
+            : model.output_s;
+    GGML_ASSERT(head_w && "GLM4_MOE MTP: missing LM head (nextn.shared_head_head or model.output)");
+
+    cur = build_lora_mm(head_w, cur, head_s);
+    cb(cur, "result_output", -1);
+
+    res->t_logits = cur;
+    ggml_build_forward_expand(gf, cur);
+}
+
 llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
@@ -154,8 +318,7 @@ llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_pa
 
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
-    // Only process up to last layer (skip final NextN layer)
-    // Final layer tensors are loaded but not processed in forward pass
+    // NextN layers are processed by graph_mtp.
     for (int il = 0; il < n_layer; ++il) {
         ggml_tensor * inpSA = inpL;
 
@@ -205,7 +368,7 @@ llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_pa
                     model.layers[il].wo, NULL, model.layers[il].wo_s,
                     Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
         }
-        if (il == n_layer - 1 && inp_out_ids) {
+        if (il == n_layer - 1 && inp_out_ids && (!cparams.embeddings_nextn || cparams.embeddings_nextn_masked)) {
             cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
             inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
         }
@@ -265,6 +428,13 @@ llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_pa
     cur = inpL;
     cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
 
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    if (cparams.embeddings_nextn && !cparams.embeddings_nextn_masked && inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
     cb(cur, "result_norm", -1);
     res->t_embd = cur;
 
diff --git a/src/models/mamba-base.cpp b/src/models/mamba-base.cpp
index 1f994ae0a..03ee3805b 100644
--- a/src/models/mamba-base.cpp
+++ b/src/models/mamba-base.cpp
@@ -182,13 +182,14 @@ ggml_tensor * llm_build_mamba_base::build_mamba2_layer(llm_graph_input_rs * inp,
     ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
     conv               = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2 * n_group * d_state, n_seqs);
 
-    // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
-    cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
-
     // d_in_proj = 2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads
 
-    // {n_embd, d_in_proj} @ {n_embd, n_seq_tokens, n_seqs} => {d_in_proj, n_seq_tokens, n_seqs}
+    // Keep the projection 2D: with a {n_embd, 1, n_seqs} batch the CUDA backend
+    // dispatches a column-batched GEMV for what is a large dense GEMM.
+    // {n_embd, d_in_proj} @ {n_embd, n_tokens} => {d_in_proj, n_tokens}
     ggml_tensor * zxBCdt = build_lora_mm(model.layers[il].ssm_in, cur, model.layers[il].ssm_in_s);
+    // {d_in_proj, n_tokens} => {d_in_proj, n_seq_tokens, n_seqs}
+    zxBCdt = ggml_reshape_3d(ctx0, zxBCdt, zxBCdt->ne[0], n_seq_tokens, n_seqs);
 
     // split the above in three
     ggml_tensor * z   = ggml_view_4d(ctx0, zxBCdt, head_dim, n_head, n_seq_tokens, n_seqs, head_dim * zxBCdt->nb[0],
@@ -290,15 +291,12 @@ ggml_tensor * llm_build_mamba_base::build_mamba2_layer(llm_graph_input_rs * inp,
             y = build_norm(y, model.layers[il].ssm_norm, NULL, LLM_NORM_RMS, il);
         }
 
-        y = ggml_reshape_3d(ctx0, y, d_inner, n_seq_tokens, n_seqs);
+        y = ggml_reshape_2d(ctx0, y, d_inner, n_seq_tokens * n_seqs);
 
-        // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
+        // {d_inner, n_embd} @ {d_inner, n_tokens} => {n_embd, n_tokens}
         cur = build_lora_mm(model.layers[il].ssm_out, y, model.layers[il].ssm_out_s);
     }
 
-    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
-    cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);
     cb(cur, "mamba_out", il);
-
     return cur;
 }
diff --git a/src/models/models.h b/src/models/models.h
index 157b05dc0..969429e3b 100644
--- a/src/models/models.h
+++ b/src/models/models.h
@@ -1412,6 +1412,10 @@ struct llama_model_glm4_moe : public llama_model_base {
         graph(const llama_model & model, const llm_graph_params & params);
     };
 
+    struct graph_mtp : public llm_graph_context {
+        graph_mtp(const llama_model & model, const llm_graph_params & params);
+    };
+
     std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
 };
 
diff --git a/tools/parser/template-analysis.cpp b/tests/test-chat-analysis.cpp
similarity index 97%
rename from tools/parser/template-analysis.cpp
rename to tests/test-chat-analysis.cpp
index bf898a229..42ad7a072 100644
--- a/tools/parser/template-analysis.cpp
+++ b/tests/test-chat-analysis.cpp
@@ -11,9 +11,9 @@
 #include <vector>
 #include <algorithm>
 
-#include "nlohmann/json.hpp"
+#include "json.h"
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 // ANSI color codes - using 256-color palette for brighter colors (all bold)
 #define ANSI_RESET       "\033[0m"
@@ -84,11 +84,12 @@ static std::string read_file(const std::string & path) {
 }
 
 static void print_usage(const char * program_name) {
-    LOG_ERR("Usage: %s [options]\n", program_name);
+    LOG_ERR("Debug the auto-parser's differential analysis: render a template with/without tools, reasoning, etc. and show the diffs.\n");
+    LOG_ERR("\nUsage: %s [options]\n", program_name);
     LOG_ERR("\nOptions:\n");
     LOG_ERR("  --template <name>       Analyze specific template from test suite (e.g., 'deepseek' or 'DeepSeek-V3.1')\n");
     LOG_ERR("  --template-file <path>  Analyze custom template file\n");
-    LOG_ERR("  --all                   Analyze all templates from test suite\n");
+    LOG_ERR("  --all                   Analyze all templates from test suite (default when no arguments are given)\n");
     LOG_ERR("\nExamples:\n");
     LOG_ERR("  %s --all\n", program_name);
     LOG_ERR("  %s --template deepseek\n", program_name);
@@ -97,14 +98,17 @@ static void print_usage(const char * program_name) {
 
 static bool parse_options(int argc, char ** argv, analysis_options & opts) {
     if (argc < 2) {
-        print_usage(argv[0]);
-        return false;
+        // default mode: analyze all templates from the test suite
+        opts.analyze_all = true;
     }
 
     for (int i = 1; i < argc; ++i) {
         std::string arg = argv[i];
 
-        if (arg == "--all") {
+        if (arg == "-h" || arg == "--help") {
+            print_usage(argv[0]);
+            return false;
+        } else if (arg == "--all") {
             opts.analyze_all = true;
         } else if (arg == "--template") {
             if (i + 1 >= argc) {
diff --git a/tools/cli/cli-context.cpp b/tools/cli/cli-context.cpp
index 3d801b73d..aa4eb7679 100644
--- a/tools/cli/cli-context.cpp
+++ b/tools/cli/cli-context.cpp
@@ -6,8 +6,7 @@
 #include "log.h"
 #include "console.h"
 
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <algorithm>
 #include <cctype>
@@ -16,7 +15,7 @@
 #include <map>
 #include <set>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 struct cli_context_impl {
     json messages      = json::array();
@@ -73,7 +72,7 @@ static std::string format_error_message(const json & err) {
 
 // err is the raw response body of a failed request; it may or may not be JSON
 static std::string format_error_message(const std::string & err) {
-    json parsed = json::parse(err, nullptr, false);
+    json parsed = json::parse_no_throw(err);
     if (!parsed.is_discarded()) {
         return format_error_message(parsed);
     }
@@ -157,7 +156,7 @@ bool cli_context::init() {
             if (!list_and_ask_models()) {
                 return false;
             }
-        } catch (const json::parse_error & e) {
+        } catch (const common_json_error & e) {
             ui::show_error(e.what());
             ui::show_message("This might be caused by an incorrect server-base endpoint URL");
             return false;
@@ -364,7 +363,7 @@ bool cli_context::generate_completion(generated_content & content_out, cli_timin
     ui::assistant_turn a;
 
     std::string err = client.post_sse("/v1/chat/completions", body.dump(), should_stop, [&](const std::string & payload) {
-        json chunk = json::parse(payload, nullptr, false);
+        json chunk = json::parse_no_throw(payload);
         if (chunk.is_discarded()) {
             return;
         }
diff --git a/tools/mtmd/clip-graph.h b/tools/mtmd/clip-graph.h
index 2cf1b683a..bbee35bea 100644
--- a/tools/mtmd/clip-graph.h
+++ b/tools/mtmd/clip-graph.h
@@ -137,9 +137,15 @@ struct clip_graph {
             int il,
             ggml_tensor * sinks = nullptr) const;
 
-    // implementation of the 2D RoPE without adding a new op in ggml
-    // this is not efficient (use double the memory), but works on all backends
-    // TODO: there was a more efficient which relies on ggml_view and ggml_rope_ext_inplace, but the rope inplace does not work well with non-contiguous tensors ; we should fix that and revert back to the original implementation in https://github.com/ggml-org/llama.cpp/pull/13065
+    // implementation of the 2D RoPE using two ggml_rope_ext calls
+    //
+    // unlike GGML_ROPE_TYPE_VISION which forces NEOX ordering, this rotates adjacent pairs (normal ordering)
+    //
+    // example:
+    //  given a single head with size = 8 --> [00000000]
+    //  dims [0, 4) rotate with pos_a, dims [4, 8) rotate with pos_b --> [aaaabbbb]
+    //  interleave_freq = false --> both halves use the same inv_freq set (like GGML_ROPE_TYPE_VISION)
+    //  interleave_freq = true  --> first half uses even inv_freq, second half uses odd inv_freq (used by pixtral)
     ggml_tensor * build_rope_2d(
         ggml_context * ctx0,
         ggml_tensor * cur,
diff --git a/tools/mtmd/clip-model.h b/tools/mtmd/clip-model.h
index fcdabd633..060938d86 100644
--- a/tools/mtmd/clip-model.h
+++ b/tools/mtmd/clip-model.h
@@ -29,10 +29,10 @@ enum patch_merge_type {
     PATCH_MERGE_SPATIAL_UNPAD,
 };
 
+// all algos are Pillow-compatible (matching PIL.Image.resize output)
 enum resize_algo {
-    RESIZE_ALGO_BILINEAR, // stretch to target resolution
-    RESIZE_ALGO_BICUBIC, // center-crop when aspect ratio doesn't match
-    RESIZE_ALGO_BICUBIC_PILLOW,
+    RESIZE_ALGO_BILINEAR,
+    RESIZE_ALGO_BICUBIC,
     RESIZE_ALGO_LANCZOS,
 };
 
@@ -73,7 +73,7 @@ struct clip_hparams {
     int32_t preproc_max_tiles = 0;
     int32_t preproc_tile_size = 0; // local tile size (deepseek-ocr)
     resize_algo image_resize_algo_rf = RESIZE_ALGO_BICUBIC;
-    resize_algo image_resize_algo_ov = RESIZE_ALGO_BILINEAR;
+    resize_algo image_resize_algo_ov = RESIZE_ALGO_BICUBIC;
     pad_style image_pad_rf = PAD_CEIL;  // padding style for the refined image (e.g. llava-1.6)
     pad_style image_pad_ov = PAD_NONE;  // padding style for the overview image (e.g. llava-1.6)
     std::array<uint8_t, 3> image_pad_color_rf = {0, 0, 0}; // padding color for refined image
diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index d4a21857a..dcd483f5e 100644
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -877,8 +877,6 @@ ggml_tensor * clip_graph::build_attn(
 }
 
 // implementation of the 2D RoPE without adding a new op in ggml
-// this is not efficient (use double the memory), but works on all backends
-// TODO: there was a more efficient which relies on ggml_view and ggml_rope_ext_inplace, but the rope inplace does not work well with non-contiguous tensors ; we should fix that and revert back to the original implementation in https://github.com/ggml-org/llama.cpp/pull/13065
 ggml_tensor * clip_graph::build_rope_2d(
     ggml_context * ctx0,
     ggml_tensor * cur,
@@ -887,9 +885,7 @@ ggml_tensor * clip_graph::build_rope_2d(
     const float freq_base,
     const bool interleave_freq
 ) {
-    const int64_t n_dim  = cur->ne[0];
-    const int64_t n_head = cur->ne[1];
-    const int64_t n_pos  = cur->ne[2];
+    const int64_t n_dim = cur->ne[0];
 
     // for example, if we have cur tensor of shape (n_dim=8, n_head, n_pos)
     // we will have a list of 4 inv_freq: 1e-0, 1e-1, 1e-2, 1e-3
@@ -903,46 +899,30 @@ ggml_tensor * clip_graph::build_rope_2d(
                                 ? std::pow(freq_base, (float)-2/n_dim)
                                 : 1.0;
 
-    // first half
-    ggml_tensor * first;
-    {
-        first = ggml_view_3d(ctx0, cur,
-            n_dim/2, n_head, n_pos,
-            cur->nb[1],
-            cur->nb[2],
-            0);
-        first = ggml_rope_ext(
-            ctx0,
-            first,
-            pos_a,      // positions
-            nullptr,    // freq factors
-            n_dim/2,    // n_dims
-            0, 0, freq_base,
-            1.0f, 0.0f, 1.0f, 0.0f, 0.0f
-        );
-    }
+    // first half, dims [0, n_dim/2)
+    cur = ggml_rope_ext(
+        ctx0,
+        cur,
+        pos_a,      // positions
+        nullptr,    // freq factors
+        n_dim/2,    // n_dims
+        0, 0, freq_base,
+        1.0f, 0.0f, 1.0f, 0.0f, 0.0f
+    );
 
-    // second half
-    ggml_tensor * second;
-    {
-        second = ggml_view_3d(ctx0, cur,
-            n_dim/2, n_head, n_pos,
-            cur->nb[1],
-            cur->nb[2],
-            n_dim/2 * ggml_element_size(cur));
-        second = ggml_rope_ext(
-            ctx0,
-            second,
-            pos_b,      // positions
-            nullptr,    // freq factors
-            n_dim/2,    // n_dims
-            0, 0, freq_base,
-            freq_scale_odd,
-            0.0f, 1.0f, 0.0f, 0.0f
-        );
-    }
+    // second half, dims [n_dim/2, n_dim)
+    cur = ggml_rope_ext(
+        ctx0,
+        cur,
+        pos_b,      // positions
+        nullptr,    // freq factors
+        n_dim/2,    // n_dims
+        0, 0, freq_base,
+        freq_scale_odd,
+        0.0f, 1.0f, 0.0f, 0.0f
+    );
+    cur = ggml_rope_set_offset(cur, n_dim/2);
 
-    cur = ggml_concat(ctx0, first, second, 0);
     return cur;
 }
 
@@ -1521,20 +1501,18 @@ struct clip_model_loader {
                         hparams.image_pad_color     = {122, 116, 104};
                         if (!hparams.image_res_candidates.empty()) {
                             hparams.image_resize_pad  = PAD_CEIL;
-                            hparams.image_resize_algo = RESIZE_ALGO_BILINEAR;
+                            hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                         } else {
                             // llava-1.6 default params
                             hparams.image_pad_ov         = PAD_NONE;
                             hparams.image_pad_rf         = PAD_CEIL;
                             hparams.image_pad_color_rf   = {122, 116, 104};
-                            hparams.image_resize_algo_rf = RESIZE_ALGO_BICUBIC;
-                            hparams.image_resize_algo_ov = RESIZE_ALGO_BILINEAR;
                         }
                     } break;
                 case PROJECTOR_TYPE_GLM_EDGE:
                     {
                         hparams.image_resize_pad  = PAD_CEIL;
-                        hparams.image_resize_algo = RESIZE_ALGO_BILINEAR;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                     } break;
                 case PROJECTOR_TYPE_MINICPMV:
                     {
@@ -1591,6 +1569,7 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_IDEFICS3:
                     {
                         // use default llava-uhd preprocessing params
+                        hparams.image_resize_algo = RESIZE_ALGO_LANCZOS;
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                         get_u32(KEY_PREPROC_IMAGE_SIZE, hparams.image_longest_edge, false);
                         hparams.set_limit_image_tokens();
@@ -1617,7 +1596,7 @@ struct clip_model_loader {
                         // ref: https://huggingface.co/mistral-community/pixtral-12b/blob/main/preprocessor_config.json
                         // TODO: verify the image_min_tokens
                         hparams.n_merge = 1; // the original pixtral does not use patch merging
-                        hparams.image_resize_algo = RESIZE_ALGO_BILINEAR;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                         hparams.rope_theta = 10000.0f;
                         get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
                         hparams.set_limit_image_tokens(8, 1024);
@@ -1645,7 +1624,7 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_DOTS3NOTE_V:
                     {
                         hparams.rope_theta = 10000.0f;
-                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC_PILLOW;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                         get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge);
                         get_u32(KEY_IMAGE_MIN_PIXELS, hparams.image_min_pixels);
                         get_u32(KEY_IMAGE_MAX_PIXELS, hparams.image_max_pixels);
@@ -1663,7 +1642,7 @@ struct clip_model_loader {
                     } break;
                 case PROJECTOR_TYPE_KIMIVL:
                     {
-                        hparams.image_resize_algo = RESIZE_ALGO_BILINEAR;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                         hparams.rope_theta = 10000.0f;
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                         // TODO: check kimivl preprocessor for exact values
@@ -1702,7 +1681,7 @@ struct clip_model_loader {
                     {
                         hparams.rope_theta = 100.0f;
                         hparams.n_merge = 3; // pooling_kernel_size
-                        hparams.image_resize_algo = RESIZE_ALGO_BILINEAR;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                         if (model.proj_type == PROJECTOR_TYPE_GEMMA4UV) {
                             // for "unified" variant, we directly use a bigger patch size, because the "token merging" is done directly on conv layer
@@ -1719,6 +1698,7 @@ struct clip_model_loader {
                         // Gemma3n uses MobileNetV5 which produces 256 tokens (16x16)
                         // Similar configuration to Gemma3
                         hparams.n_merge = 1;  // MobileNetV5 handles resizing internally
+                        hparams.image_resize_algo = RESIZE_ALGO_BILINEAR;
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                     } break;
                 case PROJECTOR_TYPE_QWEN2VL:
@@ -1726,7 +1706,7 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_QWEN3VL:
                     {
                         hparams.n_merge = 2; // default value for Qwen 2 and 2.5
-                        hparams.image_resize_algo = RESIZE_ALGO_BILINEAR;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                         get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
                         get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern, model.proj_type == PROJECTOR_TYPE_QWEN25VL); // only 2.5 requires it
                         // ref: https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/blob/main/preprocessor_config.json
@@ -1747,7 +1727,7 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_MINIMAX_M3:
                     {
                         hparams.n_merge = 2; // spatial_merge_size
-                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC_PILLOW;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                         hparams.image_resize_pad  = PAD_NONE;
                         get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
                         // n_merge is used as a divisor in clip_image_batch_encode
@@ -1772,7 +1752,7 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_MIMOVL:
                     {
                         hparams.n_merge = 2; // spatial_merge_size
-                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC_PILLOW;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                         get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
                         get_u32(string_format(KEY_N_HEAD_KV, "vision"), hparams.n_head_kv);
                         // 1D banded sliding-window radius (visual_token_window_size); required
@@ -1819,15 +1799,15 @@ struct clip_model_loader {
                         log_ffn_op = "gelu_erf";
                         hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
 
-                        // reka model performs better when using resize_bicubic, which stretches
-                        // the image to fit fixed square size
+                        // reka model performs better when the image is stretched to fit
+                        // fixed square size (no padding)
                         hparams.image_resize_pad = PAD_NONE;
                     } break;
                 case PROJECTOR_TYPE_GLM4V:
                     {
                         hparams.rope_theta = 10000.0f;
                         hparams.n_merge = 2; // default value for GLM4-V
-                        hparams.image_resize_algo = RESIZE_ALGO_BILINEAR;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                         get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
                         hparams.set_limit_image_tokens(8, 4096);
                         hparams.set_warmup_n_tokens(46*46); // avoid OOM on warmup
@@ -1835,6 +1815,7 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_LLAMA4:
                     {
                         hparams.rope_theta = 10000.0f;
+                        hparams.image_resize_algo = RESIZE_ALGO_BILINEAR;
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                         set_llava_uhd_res_candidates(model, 3);
                     } break;
@@ -1946,7 +1927,7 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_PADDLEOCR:
                     {
                         hparams.n_merge = 2;
-                        hparams.image_resize_algo = RESIZE_ALGO_BILINEAR;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                         get_u32(KEY_IMAGE_MIN_PIXELS, hparams.image_min_pixels);
                         get_u32(KEY_IMAGE_MAX_PIXELS, hparams.image_max_pixels);
 
@@ -1958,7 +1939,7 @@ struct clip_model_loader {
                         hparams.patch_size = 16;
                         hparams.image_size = 1024;
                         hparams.warmup_image_size = 1024;
-                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC_PILLOW;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                         hparams.image_pad_color = {127, 127, 127};
 
                         get_u32(KEY_SAM_N_BLOCK, hparams.sam_n_layer, true);
@@ -1988,7 +1969,7 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_HUNYUANVL:
                     {
                         hparams.n_merge = 2;
-                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC_PILLOW;
+                        hparams.image_resize_algo = RESIZE_ALGO_LANCZOS;
                         hparams.image_resize_pad = PAD_NONE;
                         hparams.ffn_op = FFN_GELU;
                         hparams.set_limit_image_tokens(256, 16384);
@@ -2061,12 +2042,12 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_JANUS_PRO:
                     {
                         hparams.image_pad_color   = {127, 127, 127};
-                        hparams.image_resize_algo = RESIZE_ALGO_BILINEAR;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                     } break;
                 case PROJECTOR_TYPE_GRANITE4_VISION:
                     {
                         // SigLIP tower.
-                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC_PILLOW;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
                         hparams.image_resize_pad = PAD_CEIL;
 
                         // NOTE: feature_layers loaded in common path as optional
diff --git a/tools/mtmd/models/gemma4v.cpp b/tools/mtmd/models/gemma4v.cpp
index 87cbd43fc..448438947 100644
--- a/tools/mtmd/models/gemma4v.cpp
+++ b/tools/mtmd/models/gemma4v.cpp
@@ -44,51 +44,31 @@ ggml_cgraph * clip_graph_gemma4v::build() {
 
     // similar to build_rope_2d, but use neox ordering
     auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
-        const int64_t n_dim  = cur->ne[0];
-        const int64_t n_head = cur->ne[1];
-        const int64_t n_pos  = cur->ne[2];
+        const int64_t n_dim = cur->ne[0];
 
-        // first half
-        ggml_tensor * first;
-        {
-            first = ggml_view_4d(ctx0, cur,
-                n_dim/2, n_head, n_pos, n_batch,
-                cur->nb[1],
-                cur->nb[2],
-                cur->nb[3],
-                0);
-            first = ggml_rope_ext(
-                ctx0,
-                first,
-                pos_x,      // positions
-                nullptr,    // freq factors
-                n_dim/2,    // n_dims
-                GGML_ROPE_TYPE_NEOX, 0, hparams.rope_theta,
-                1.0f, 0.0f, 1.0f, 0.0f, 0.0f
-            );
-        }
+        // first half, dims [0, n_dim/2)
+        cur = ggml_rope_ext(
+            ctx0,
+            cur,
+            pos_x,      // positions
+            nullptr,    // freq factors
+            n_dim/2,    // n_dims
+            GGML_ROPE_TYPE_NEOX, 0, hparams.rope_theta,
+            1.0f, 0.0f, 1.0f, 0.0f, 0.0f
+        );
 
-        // second half
-        ggml_tensor * second;
-        {
-            second = ggml_view_4d(ctx0, cur,
-                n_dim/2, n_head, n_pos, n_batch,
-                cur->nb[1],
-                cur->nb[2],
-                cur->nb[3],
-                n_dim/2 * ggml_element_size(cur));
-            second = ggml_rope_ext(
-                ctx0,
-                second,
-                pos_y,      // positions
-                nullptr,    // freq factors
-                n_dim/2,    // n_dims
-                GGML_ROPE_TYPE_NEOX, 0, hparams.rope_theta,
-                1.0f, 0.0f, 1.0f, 0.0f, 0.0f
-            );
-        }
+        // second half, dims [n_dim/2, n_dim)
+        cur = ggml_rope_ext(
+            ctx0,
+            cur,
+            pos_y,      // positions
+            nullptr,    // freq factors
+            n_dim/2,    // n_dims
+            GGML_ROPE_TYPE_NEOX, 0, hparams.rope_theta,
+            1.0f, 0.0f, 1.0f, 0.0f, 0.0f
+        );
+        cur = ggml_rope_set_offset(cur, n_dim/2);
 
-        cur = ggml_concat(ctx0, first, second, 0);
         return cur;
     };
 
diff --git a/tools/mtmd/models/minimax-m3.cpp b/tools/mtmd/models/minimax-m3.cpp
index 447621754..256e53105 100644
--- a/tools/mtmd/models/minimax-m3.cpp
+++ b/tools/mtmd/models/minimax-m3.cpp
@@ -2,30 +2,22 @@
 
 ggml_tensor * clip_graph_minimax_m3::apply_rope(
         ggml_tensor * x, ggml_tensor * pos_h, ggml_tensor * pos_w) {
-    const int64_t Hn  = x->ne[1];
-    const int64_t P   = x->ne[2];
-    const size_t  es  = ggml_element_size(x);
-    const int     dh  = (int) x->ne[0];
-    const int     axd = 2 * ((2 * (dh / 2) / 3) / 2);
+    const int dh  = (int) x->ne[0];
+    const int axd = 2 * ((2 * (dh / 2) / 3) / 2);
 
-    GGML_ASSERT(x->nb[0] == es);
     GGML_ASSERT(3 * axd <= dh);
 
     const float th  = hparams.rope_theta;
 
     // layout of x is [t, h, w, pad]
     // t is unrotated, h and w are rotated, pad is unrotated
-    // note: everything from n_dims onward untouched, so w and pad are rotated in one call.
-    auto sl = [&](int off, int n) {
-        return ggml_cont(ctx0, ggml_view_3d(ctx0, x, n, Hn, P, x->nb[1], x->nb[2], (size_t) off * es));
-    };
-    ggml_tensor * t = sl(0,       axd);
-    ggml_tensor * h = sl(axd,     axd);
-    ggml_tensor * w = sl(2 * axd, dh - 2 * axd); // w + pad
+    x = ggml_rope_ext(ctx0, x, pos_h, nullptr, axd, GGML_ROPE_TYPE_NEOX, 0, th, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+    x = ggml_rope_set_offset(x, axd);
 
-    h = ggml_rope_ext(ctx0, h, pos_h, nullptr, axd, GGML_ROPE_TYPE_NEOX, 0, th, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
-    w = ggml_rope_ext(ctx0, w, pos_w, nullptr, axd, GGML_ROPE_TYPE_NEOX, 0, th, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
-    return ggml_concat(ctx0, ggml_concat(ctx0, t, h, 0), w, 0);
+    x = ggml_rope_ext(ctx0, x, pos_w, nullptr, axd, GGML_ROPE_TYPE_NEOX, 0, th, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+    x = ggml_rope_set_offset(x, 2 * axd);
+
+    return x;
 }
 
 ggml_cgraph * clip_graph_minimax_m3::build() {
diff --git a/tools/mtmd/mtmd-helper.cpp b/tools/mtmd/mtmd-helper.cpp
index 412b781df..42c52859f 100644
--- a/tools/mtmd/mtmd-helper.cpp
+++ b/tools/mtmd/mtmd-helper.cpp
@@ -43,6 +43,11 @@
 #ifdef MTMD_VIDEO
 #include "sheredom/subprocess.h"
 #include <thread>
+#ifndef _WIN32
+#include <csignal>
+#include <fcntl.h>
+#include <pthread.h>
+#endif
 #endif
 
 //
@@ -523,7 +528,8 @@ struct mtmd_helper_video {
     // RAII wrapper for managing subprocess
     struct subprocess_handle {
         struct subprocess_s proc = {};
-        bool alive = false;
+        bool created = false; // process exists and must be cleaned up
+        bool alive   = false; // process can still give us data
         std::thread feeder;
 
         subprocess_handle() = default;
@@ -532,18 +538,27 @@ struct mtmd_helper_video {
         ~subprocess_handle() { stop(); }
 
         void stop() {
-            if (alive) {
-                subprocess_terminate(&proc);
+            // note: alive becomes false on stdout EOF, but the process still needs cleanup
+            if (!created) {
+                return;
             }
+            subprocess_terminate(&proc);
+#ifdef _WIN32
+            // no SIGPIPE on windows: a blocked feeder only gets a broken pipe once we close our read end of the child stdin
+            if (proc.hStdInput) {
+                CloseHandle(proc.hStdInput);
+                proc.hStdInput = nullptr;
+            }
+#endif
             // join before destroy: feeder holds a FILE* from subprocess_stdin;
             // subprocess_destroy closes it, so the thread must finish first
             if (feeder.joinable()) {
                 feeder.join();
             }
-            if (alive) {
-                subprocess_destroy(&proc);
-                alive = false;
-            }
+            subprocess_join(&proc, nullptr); // reap the child, or else it stays a zombie
+            subprocess_destroy(&proc);
+            created = false;
+            alive   = false;
         }
 
         FILE * stdout_pipe() {
@@ -553,10 +568,21 @@ struct mtmd_helper_video {
         // buf is tied to lifetime of mtmd_helper_video, so it's guaranteed to outlive the feeder thread
         void start_feeder(const std::vector<uint8_t> & buf) {
             feeder = std::thread([this, &buf]() {
+#ifndef _WIN32
+                // ffmpeg can exit before it reads all the input, for example when ffprobe already got the metadata.
+                // the write below must then fail with EPIPE, instead of killing the process with SIGPIPE
+                sigset_t sigpipe_set;
+                sigemptyset(&sigpipe_set);
+                sigaddset(&sigpipe_set, SIGPIPE);
+                pthread_sigmask(SIG_BLOCK, &sigpipe_set, nullptr); // linux sends the signal to the writing thread
+#endif
                 FILE * f = subprocess_stdin(&proc);
                 if (!f) {
                     return;
                 }
+#ifdef F_SETNOSIGPIPE
+                fcntl(fileno(f), F_SETNOSIGPIPE, 1); // macos/bsd send it to the process, so turn it off per fd
+#endif
                 fwrite(buf.data(), 1, buf.size(), f);
                 fclose(f);
                 proc.stdin_file = nullptr; // prevent double-close in subprocess_destroy
@@ -602,7 +628,8 @@ struct mtmd_helper_video {
             LOG_ERR("%s: failed to launch ffprobe\n", __func__);
             return false;
         }
-        probe_sp.alive = true;
+        probe_sp.created = true;
+        probe_sp.alive   = true;
 
         if (is_buf_input()) {
             probe_sp.start_feeder(input_buf);
@@ -674,6 +701,11 @@ struct mtmd_helper_video {
         }
 
         cmd.push_back("-nostdin");
+        if (is_buf_input()) {
+            // remove the 64KB read-ahead limit of cache:, or else ffmpeg cannot reach a moov atom at end of file
+            cmd.push_back("-read_ahead_limit");
+            cmd.push_back("-1");
+        }
         cmd.push_back("-i");
         // cache:pipe:0 wraps stdin with a seekable in-memory cache, letting ffmpeg seek
         // backwards for container headers (e.g. MP4 moov atom at end of file)
@@ -712,7 +744,8 @@ struct mtmd_helper_video {
             subprocess_option_search_user_path | subprocess_option_inherit_environment,
             &sp.proc);
 
-        sp.alive = (ret == 0);
+        sp.created = (ret == 0);
+        sp.alive   = (ret == 0);
         LOG_DBG("%s: subprocess_create ret=%d proc_alive=%d\n", __func__, ret, (int)sp.alive);
 
         if (sp.alive && is_buf_input()) {
diff --git a/tools/mtmd/mtmd-image.cpp b/tools/mtmd/mtmd-image.cpp
index 0d9db4f62..0dda8770f 100644
--- a/tools/mtmd/mtmd-image.cpp
+++ b/tools/mtmd/mtmd-image.cpp
@@ -58,22 +58,7 @@ struct img_tool {
 
         if (padding == PAD_NONE) {
             // direct resize
-            switch (algo) {
-                case RESIZE_ALGO_BILINEAR:
-                    resize_bilinear(src, dst, target_resolution.width, target_resolution.height);
-                    break;
-                case RESIZE_ALGO_BICUBIC:
-                    resize_bicubic(src, dst, target_resolution.width, target_resolution.height);
-                    break;
-                case RESIZE_ALGO_BICUBIC_PILLOW:
-                    resize_bicubic_pillow(src, dst, target_resolution.width, target_resolution.height);
-                    break;
-                case RESIZE_ALGO_LANCZOS:
-                    resize_lanczos_pillow(src, dst, target_resolution.width, target_resolution.height);
-                    break;
-                default:
-                    throw std::runtime_error("Unsupported resize algorithm");
-            }
+            resize_pillow(src, dst, target_resolution.width, target_resolution.height, algo);
         } else {
             // resize with padding
             clip_image_u8 resized_image;
@@ -90,22 +75,7 @@ struct img_tool {
                 new_height = std::min(static_cast<int>(std::ceil(src.get_size().height * scale)), target_resolution.height);
             }
 
-            switch (algo) {
-                case RESIZE_ALGO_BILINEAR:
-                    resize_bilinear(src, resized_image, new_width, new_height);
-                    break;
-                case RESIZE_ALGO_BICUBIC:
-                    resize_bicubic(src, resized_image, new_width, new_height);
-                    break;
-                case RESIZE_ALGO_BICUBIC_PILLOW:
-                    resize_bicubic_pillow(src, resized_image, new_width, new_height);
-                    break;
-                case RESIZE_ALGO_LANCZOS:
-                    resize_lanczos_pillow(src, resized_image, new_width, new_height);
-                    break;
-                default:
-                    throw std::runtime_error("Unsupported resize algorithm");
-            }
+            resize_pillow(src, resized_image, new_width, new_height, algo);
 
             // fill dst with pad_color
             fill(dst, pad_color);
@@ -224,152 +194,37 @@ struct img_tool {
     }
 
 private:
-    // Bilinear resize function
-    static void resize_bilinear(const clip_image_u8 & src, clip_image_u8 & dst, int target_width, int target_height) {
-        const auto src_size = src.get_size();
-        if (src_size.width == 0 || src_size.height == 0) { dst.set_size({0, 0}, false); return; }
-        if (target_width  <= 0) target_width  = 1;
-        if (target_height <= 0) target_height = 1;
-
-        dst.set_size({target_width, target_height}, false);
-
-        if (src.is_placeholder()) {
-            // no-op for placeholder image, just set the size and return
-            return;
-        }
-
-        float x_ratio = target_width  > 1 ? static_cast<float>(src_size.width  - 1) / (target_width  - 1) : 0.0f;
-        float y_ratio = target_height > 1 ? static_cast<float>(src_size.height - 1) / (target_height - 1) : 0.0f;
-
-        for (int y = 0; y < target_height; ++y) {
-            for (int x = 0; x < target_width; ++x) {
-                float px = x * x_ratio;
-                float py = y * y_ratio;
-
-                int x0 = std::min(static_cast<int>(px), src_size.width  - 1);
-                int y0 = std::min(static_cast<int>(py), src_size.height - 1);
-                int x1 = std::min(x0 + 1, src_size.width  - 1);
-                int y1 = std::min(y0 + 1, src_size.height - 1);
-
-                float xf = px - x0;
-                float yf = py - y0;
-
-                const auto p00 = src.get_pixel(x0, y0);
-                const auto p10 = src.get_pixel(x1, y0);
-                const auto p01 = src.get_pixel(x0, y1);
-                const auto p11 = src.get_pixel(x1, y1);
-
-                std::array<uint8_t, 3> pixel;
-                for (int c = 0; c < 3; ++c) {
-                    float top    = lerp(static_cast<float>(p00[c]), static_cast<float>(p10[c]), xf);
-                    float bottom = lerp(static_cast<float>(p01[c]), static_cast<float>(p11[c]), xf);
-                    pixel[c] = static_cast<uint8_t>(lerp(top, bottom, yf));
-                }
-                dst.set_pixel(x, y, pixel);
-            }
-        }
-    }
-
-    // Bicubic resize function
-    // part of image will be cropped if the aspect ratio is different
-    static void resize_bicubic(const clip_image_u8 & img, clip_image_u8 & dst, int target_width, int target_height) {
-        const auto img_size = img.get_size();
-        const int nx = img_size.width;
-        const int ny = img_size.height;
-
-        dst.set_size({target_width, target_height}, false);
-
-        if (img.is_placeholder()) {
-            // no-op for placeholder image, just set the size and return
-            return;
-        }
-
-        float Cc;
-        float C[5] = {};
-        float d0, d2, d3, a0, a1, a2, a3;
-        int i, j, k, jj;
-        int x, y;
-        float dx, dy;
-        float tx, ty;
-
-        tx = (float)nx / (float)target_width;
-        ty = (float)ny / (float)target_height;
-
-        // Bicubic interpolation; adapted from ViT.cpp, inspired from :
-        //    -> https://github.com/yglukhov/bicubic-interpolation-image-processing/blob/master/libimage.c#L36
-        //    -> https://en.wikipedia.org/wiki/Bicubic_interpolation
-
-        for (i = 0; i < target_height; i++) {
-            for (j = 0; j < target_width; j++) {
-                x = (int)(tx * j);
-                y = (int)(ty * i);
-
-                dx = tx * j - x;
-                dy = ty * i - y;
-
-                std::array<uint8_t, 3> pixel;
-                for (k = 0; k < 3; k++) {
-                    for (jj = 0; jj <= 3; jj++) {
-                        d0 = img.get_pixel(clip(x - 1, 0, nx - 1), clip(y - 1 + jj, 0, ny - 1))[k] - img.get_pixel(clip(x, 0, nx - 1), clip(y - 1 + jj, 0, ny - 1))[k];
-                        d2 = img.get_pixel(clip(x + 1, 0, nx - 1), clip(y - 1 + jj, 0, ny - 1))[k] - img.get_pixel(clip(x, 0, nx - 1), clip(y - 1 + jj, 0, ny - 1))[k];
-                        d3 = img.get_pixel(clip(x + 2, 0, nx - 1), clip(y - 1 + jj, 0, ny - 1))[k] - img.get_pixel(clip(x, 0, nx - 1), clip(y - 1 + jj, 0, ny - 1))[k];
-                        a0 = img.get_pixel(clip(x, 0, nx - 1), clip(y - 1 + jj, 0, ny - 1))[k];
-
-                        a1 = -1.0 / 3 * d0 + d2 - 1.0 / 6 * d3;
-                        a2 =  1.0 / 2 * d0 +      1.0 / 2 * d2;
-                        a3 = -1.0 / 6 * d0 -      1.0 / 2 * d2 + 1.0 / 6 * d3;
-
-                        C[jj] = a0 + a1 * dx + a2 * dx * dx + a3 * dx * dx * dx;
-
-                        d0 = C[0] - C[1];
-                        d2 = C[2] - C[1];
-                        d3 = C[3] - C[1];
-                        a0 = C[1];
-                        a1 = -1.0 / 3 * d0 + d2 - 1.0 / 6 * d3;
-                        a2 =  1.0 / 2 * d0 +      1.0 / 2 * d2;
-                        a3 = -1.0 / 6 * d0 -      1.0 / 2 * d2 + 1.0 / 6 * d3;
-                        Cc = a0 + a1 * dy + a2 * dy * dy + a3 * dy * dy * dy;
-
-                        const uint8_t Cc2 = std::min(std::max(std::round(Cc), 0.0f), 255.0f);
-                        pixel[k] = Cc2;
-                    }
-                }
-                dst.set_pixel(j, i, pixel);
-            }
-        }
-    }
-
-    // Pillow-compatible separable resampling (Bicubic and Lanczos)
+    // Pillow-compatible separable resampling (Bilinear, Bicubic and Lanczos)
     // Adapted from https://github.com/python-pillow/Pillow/blob/main/src/libImaging/Resample.c
     //
     // Key properties:
     // 1. Separable filtering: horizontal pass followed by vertical pass
     // 2. Pre-computes normalized filter coefficients for each output pixel
     // 3. Fixed-point integer arithmetic (22 fractional bits) for speed and determinism
-    static bool resize_bicubic_pillow(const clip_image_u8 & img, clip_image_u8 & dst, int target_width, int target_height) {
-        return resize_pillow(img, dst, target_width, target_height, /*use_lanczos=*/false);
-    }
-
-    // Lanczos-3 (support radius 3), matches Pillow's Image.LANCZOS
-    static bool resize_lanczos_pillow(const clip_image_u8 & img, clip_image_u8 & dst, int target_width, int target_height) {
-        return resize_pillow(img, dst, target_width, target_height, /*use_lanczos=*/true);
-    }
-
     static bool resize_pillow(
             const clip_image_u8 & img,
             clip_image_u8 & dst,
             int target_width,
             int target_height,
-            bool use_lanczos) {
+            resize_algo algo) {
         // Fixed-point precision: 22 bits = 32 (int32_t) - 8 (uint8_t pixels) - 2 (headroom for accumulation)
         // This allows encoding fractional weights as integers: weight * 2^22
         const int PRECISION_BITS = 32 - 8 - 2;
 
-        // Resample filter: Lanczos-3 (support [-3, 3]) or bicubic with a = -0.5 (support [-2, 2])
-        // Note: GGML/PyTorch bicubic uses a = -0.75, Pillow uses a = -0.5
+        // Filter support radius
+        double filter_support;
+        switch (algo) {
+            case RESIZE_ALGO_BILINEAR: filter_support = 1.0; break;
+            case RESIZE_ALGO_BICUBIC:  filter_support = 2.0; break;
+            case RESIZE_ALGO_LANCZOS:  filter_support = 3.0; break;
+            default:
+                throw std::runtime_error("Unsupported resize algorithm");
+        }
+
         // Returns filter weight for distance x from pixel center
-        auto resample_filter = [use_lanczos](double x) -> double {
-            if (use_lanczos) {
+        // Note: for bicubic, Pillow uses a = -0.5 while GGML/PyTorch use a = -0.75
+        auto resample_filter = [algo](double x) -> double {
+            if (algo == RESIZE_ALGO_LANCZOS) {
                 if (-3.0 <= x && x < 3.0) {
                     auto sinc = [](double v) {
                         if (v == 0.0) {
@@ -383,10 +238,15 @@ private:
                 return 0.0;
             }
 
-            constexpr double a = -0.5;
             if (x < 0.0) {
                 x = -x;
             }
+
+            if (algo == RESIZE_ALGO_BILINEAR) {
+                return x < 1.0 ? 1.0 - x : 0.0;
+            }
+
+            constexpr double a = -0.5;
             if (x < 1.0) {
                 return ((a + 2.0) * x - (a + 3.0)) * x * x + 1;
             }
@@ -396,9 +256,6 @@ private:
             return 0.0;  // Zero outside [-2, 2]
         };
 
-        // Filter support radius: 2 for bicubic, 3 for lanczos
-        const double filter_support = use_lanczos ? 3.0 : 2.0;
-
         // Clipping function for 8-bit values
         auto clip8 = [](int val) -> uint8_t {
             if (val < 0) return 0;
@@ -493,100 +350,92 @@ private:
             const double fxp_scale = std::ldexp(1.0, PRECISION_BITS); // 1.0 * 2^PRECISION_BITS
 
             for (int i = 0; i < outSize * ksize; i++) {
-                if (use_lanczos) {
-                    // Pillow adds +/- 0.5 then truncates toward zero; std::round would round twice
-                    const double rounded = pre_weights[i] * fxp_scale + (pre_weights[i] < 0 ? -0.5 : 0.5);
-                    weights[i] = static_cast<int32_t>(rounded);
-                    continue;
-                }
-                double tmp_val = pre_weights[i] * fxp_scale;
-                if (pre_weights[i] < 0) {
-                    tmp_val -= 0.5;
-                } else {
-                    tmp_val += 0.5;
-                }
-                tmp_val = std::round(tmp_val);
-                tmp_val = std::clamp(tmp_val,
-                                     static_cast<double>(std::numeric_limits<int32_t>::min()),
-                                     static_cast<double>(std::numeric_limits<int32_t>::max()));
-                weights[i] = static_cast<int32_t>(tmp_val);
+                // Pillow adds +/- 0.5 then truncates toward zero; std::round would round twice
+                const double rounded = pre_weights[i] * fxp_scale + (pre_weights[i] < 0 ? -0.5 : 0.5);
+                weights[i] = static_cast<int32_t>(rounded);
             }
 
             return ksize;
         };
 
         // Horizontal resampling pass
-        // Resizes width from imIn to out_nx, preserving height
-        auto resample_horizontal = [&](const clip_image_u8 & imIn, clip_image_u8 & imOut,
+        // Resizes width from src to out_nx, preserving height
+        auto resample_horizontal = [&](const uint8_t * src, int in_nx, int in_ny,
                                        int out_nx,
                                        int ksize, const std::vector<int> & bounds, const std::vector<int32_t> & weights) {
-            const int in_ny = imIn.get_size().height;
-            imOut.set_size({out_nx, in_ny}, false);
+            std::vector<uint8_t> out((size_t) out_nx * in_ny * 3);
 
             // Process each row independently
             for (int yy = 0; yy < in_ny; yy++) {
+                const uint8_t * src_row = src + (size_t) yy * in_nx * 3;
+                uint8_t * dst_row = out.data() + (size_t) yy * out_nx * 3;
+
                 // For each output pixel in this row
                 for (int xx = 0; xx < out_nx; xx++) {
-                    // Get the range of input pixels and filter coefficients
-                    int xmin = bounds[xx * 2 + 0];  // First input pixel index
-                    int xcnt = bounds[xx * 2 + 1];  // Number of input pixels
+                    const int xmin = bounds[xx * 2 + 0];  // First input pixel index
+                    const int xcnt = bounds[xx * 2 + 1];  // Number of input pixels
+                    const int32_t * k = &weights[xx * ksize];
+                    const uint8_t * p = src_row + (size_t) xmin * 3;
 
-                    // Initialize accumulators for RGB channels with rounding bias (0.5 in fixed-point)
+                    // Accumulators for RGB channels, with rounding bias (0.5 in fixed-point)
                     int32_t ss0 = 1 << (PRECISION_BITS - 1);
                     int32_t ss1 = 1 << (PRECISION_BITS - 1);
                     int32_t ss2 = 1 << (PRECISION_BITS - 1);
 
                     // Convolve: sum weighted input pixels
                     for (int x = 0; x < xcnt; x++) {
-                        const auto src_px = imIn.get_pixel(x + xmin, yy);
-                        ss0 += src_px[0] * weights[xx * ksize + x];  // R channel
-                        ss1 += src_px[1] * weights[xx * ksize + x];  // G channel
-                        ss2 += src_px[2] * weights[xx * ksize + x];  // B channel
+                        ss0 += p[0] * k[x];
+                        ss1 += p[1] * k[x];
+                        ss2 += p[2] * k[x];
+                        p += 3;
                     }
 
                     // Convert back from fixed-point (divide by 2^PRECISION_BITS) and clamp to [0,255]
-                    imOut.set_pixel(xx, yy, {clip8(ss0 >> PRECISION_BITS),
-                                             clip8(ss1 >> PRECISION_BITS),
-                                             clip8(ss2 >> PRECISION_BITS)});
+                    dst_row[xx * 3 + 0] = clip8(ss0 >> PRECISION_BITS);
+                    dst_row[xx * 3 + 1] = clip8(ss1 >> PRECISION_BITS);
+                    dst_row[xx * 3 + 2] = clip8(ss2 >> PRECISION_BITS);
                 }
             }
+
+            return out;
         };
 
         // Vertical resampling pass
-        // Resizes height from imIn to out_ny, preserving width
-        auto resample_vertical = [&](const clip_image_u8 & imIn, clip_image_u8 & imOut,
+        // Resizes height from src to out_ny, preserving width
+        // Accumulates whole rows at once (contiguous access, auto-vectorizes well)
+        auto resample_vertical = [&](const uint8_t * src, int in_nx,
                                      int out_ny,
                                      int ksize, const std::vector<int> & bounds, const std::vector<int32_t> & weight) {
-            const int in_nx = imIn.get_size().width;
-            imOut.set_size({in_nx, out_ny}, false);
+            const size_t row_elems = (size_t) in_nx * 3;
+            std::vector<uint8_t> out(row_elems * out_ny);
+            std::vector<int32_t> acc(row_elems);
 
             // For each output row
             for (int yy = 0; yy < out_ny; yy++) {
-                // Get the range of input rows and filter coefficients
-                int ymin = bounds[yy * 2 + 0];  // First input row index
-                int ycnt = bounds[yy * 2 + 1];  // Number of input rows
+                const int ymin = bounds[yy * 2 + 0];  // First input row index
+                const int ycnt = bounds[yy * 2 + 1];  // Number of input rows
+                const int32_t * k = &weight[yy * ksize];
 
-                // Process each column in this output row
-                for (int xx = 0; xx < in_nx; xx++) {
-                    // Initialize accumulators for RGB channels with rounding bias
-                    int32_t ss0 = 1 << (PRECISION_BITS - 1);
-                    int32_t ss1 = 1 << (PRECISION_BITS - 1);
-                    int32_t ss2 = 1 << (PRECISION_BITS - 1);
+                // Rounding bias (0.5 in fixed-point)
+                std::fill(acc.begin(), acc.end(), 1 << (PRECISION_BITS - 1));
 
-                    // Convolve: sum weighted input pixels vertically
-                    for (int y = 0; y < ycnt; y++) {
-                        const auto src_px = imIn.get_pixel(xx, y + ymin);
-                        ss0 += src_px[0] * weight[yy * ksize + y];  // R channel
-                        ss1 += src_px[1] * weight[yy * ksize + y];  // G channel
-                        ss2 += src_px[2] * weight[yy * ksize + y];  // B channel
+                // Convolve: accumulate each weighted input row
+                for (int y = 0; y < ycnt; y++) {
+                    const uint8_t * src_row = src + (size_t) (ymin + y) * row_elems;
+                    const int32_t w = k[y];
+                    for (size_t i = 0; i < row_elems; i++) {
+                        acc[i] += src_row[i] * w;
                     }
+                }
 
-                    // Convert back from fixed-point and clamp to [0,255]
-                    imOut.set_pixel(xx, yy, {clip8(ss0 >> PRECISION_BITS),
-                                             clip8(ss1 >> PRECISION_BITS),
-                                             clip8(ss2 >> PRECISION_BITS)});
+                // Convert back from fixed-point and clamp to [0,255]
+                uint8_t * dst_row = out.data() + (size_t) yy * row_elems;
+                for (size_t i = 0; i < row_elems; i++) {
+                    dst_row[i] = clip8(acc[i] >> PRECISION_BITS);
                 }
             }
+
+            return out;
         };
 
         // Main resampling logic using separable two-pass approach
@@ -610,36 +459,25 @@ private:
         }
 
         // Perform two-pass resampling
+        const uint8_t * src = img.get_ro_buf().data();
         if (need_horizontal && need_vertical) {
-            // Both horizontal and vertical
-            clip_image_u8 temp;
-            resample_horizontal(img, temp, target_width, ksize_horiz, bounds_horiz, weights_horiz);
-            resample_vertical(temp, dst, target_height, ksize_vert, bounds_vert, weights_vert);
+            auto temp = resample_horizontal(src, src_width, src_height, target_width, ksize_horiz, bounds_horiz, weights_horiz);
+            dst.set_size({target_width, target_height}, false);
+            dst.cpy_buf(resample_vertical(temp.data(), target_width, target_height, ksize_vert, bounds_vert, weights_vert));
         } else if (need_horizontal) {
-            // Only horizontal
-            resample_horizontal(img, dst, target_width, ksize_horiz, bounds_horiz, weights_horiz);
+            dst.set_size({target_width, src_height}, false);
+            dst.cpy_buf(resample_horizontal(src, src_width, src_height, target_width, ksize_horiz, bounds_horiz, weights_horiz));
         } else if (need_vertical) {
-            // Only vertical
-            resample_vertical(img, dst, target_height, ksize_vert, bounds_vert, weights_vert);
+            dst.set_size({src_width, target_height}, false);
+            dst.cpy_buf(resample_vertical(src, src_width, target_height, ksize_vert, bounds_vert, weights_vert));
         } else {
             // No resizing needed - direct copy
-            dst.set_size(img.get_size(), img.is_placeholder());
-            if (!img.is_placeholder()) {
-                dst.cpy_buf(img.get_ro_buf());
-            }
+            dst.set_size(img.get_size(), false);
+            dst.cpy_buf(img.get_ro_buf());
         }
 
         return true;
     }
-
-    static inline int clip(int x, int lower, int upper) {
-        return std::max(lower, std::min(x, upper));
-    }
-
-    // Linear interpolation between two points
-    static inline float lerp(float s, float e, float t) {
-        return s + (e - s) * t;
-    }
 };
 
 
@@ -1264,7 +1102,7 @@ mtmd_image_preproc_out mtmd_image_preprocessor_deepseekocr::preprocess(const cli
     clip_image_u8 padded;
     img_tool::resize(img, padded,
                     { base_size, base_size },
-                    RESIZE_ALGO_BICUBIC_PILLOW,
+                    RESIZE_ALGO_BICUBIC,
                     PAD_NEAREST,
                     hparams.image_pad_color);
     output.append_overview(hparams, padded, true);
@@ -1280,7 +1118,7 @@ mtmd_image_preproc_out mtmd_image_preprocessor_deepseekocr::preprocess(const cli
         grid_h = grid.height;
 
         clip_image_u8 refined;
-        img_tool::resize(img, refined, { tile_size * grid_w, tile_size * grid_h }, RESIZE_ALGO_BICUBIC_PILLOW,
+        img_tool::resize(img, refined, { tile_size * grid_w, tile_size * grid_h }, RESIZE_ALGO_BICUBIC,
                          PAD_NONE);
 
         for (int row = 0; row < grid_h; row++) {
diff --git a/tools/parser/CMakeLists.txt b/tools/parser/CMakeLists.txt
deleted file mode 100644
index a8df0e7e6..000000000
--- a/tools/parser/CMakeLists.txt
+++ /dev/null
@@ -1,20 +0,0 @@
-if (NOT WIN32 OR NOT BUILD_SHARED_LIBS)
-    # this tool is disabled on Windows when building with shared libraries because it uses internal functions not exported with LLAMA_API
-    set(TARGET llama-debug-template-parser)
-    add_executable(${TARGET} debug-template-parser.cpp)
-    target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
-    target_compile_features(${TARGET} PRIVATE cxx_std_17)
-
-    if(LLAMA_TOOLS_INSTALL)
-        install(TARGETS ${TARGET} RUNTIME)
-    endif()
-endif()
-
-set(TARGET llama-template-analysis)
-add_executable(${TARGET} template-analysis.cpp)
-target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
-
-if(LLAMA_TOOLS_INSTALL)
-    install(TARGETS ${TARGET} RUNTIME)
-endif()
diff --git a/tools/parser/debug-template-parser.cpp b/tools/parser/debug-template-parser.cpp
deleted file mode 100644
index 8a916f79c..000000000
--- a/tools/parser/debug-template-parser.cpp
+++ /dev/null
@@ -1,469 +0,0 @@
-#include "../src/llama-grammar.h"
-#include "chat-auto-parser.h"
-#include "chat.h"
-#include "common.h"
-#include "gguf.h"
-#include "jinja/runtime.h"
-#include "log.h"
-#include "nlohmann/json.hpp"
-#include "peg-parser.h"
-
-#include <fstream>
-#include <iterator>
-#include <numeric>
-#include <optional>
-#include <sstream>
-#include <string>
-
-using json = nlohmann::ordered_json;
-
-enum class output_mode {
-    ANALYSIS,  // Only output analysis results (default)
-    TEMPLATE,  // Only output rendered template
-    BOTH       // Output both
-};
-
-enum class input_message_type {
-    NONE,                    // Don't render any message scenarios (only analysis)
-    CONTENT_ONLY,            // Simple assistant message with content
-    REASONING_CONTENT,       // Message with reasoning_content + content
-    TOOL_CALL_ONLY,          // Message with tool_calls only
-    CONTENT_TOOL_CALL,       // Message with content + tool_calls
-    REASONING_TOOL_CALL,     // Message with reasoning_content + tool_calls
-    CONTENT_FAKE_TOOL_CALL,  // Message with content but no actual tool_calls (for testing)
-    ALL                      // Render all scenarios
-};
-
-struct debug_options {
-    std::string        template_path;
-    bool               with_tools        = true;
-    bool               generation_prompt = true;
-    bool               enable_reasoning  = true;
-    bool               debug_jinja       = false;
-    bool               force_tool_call   = false;
-    bool               parallel_tool_calls = true;
-    output_mode        mode              = output_mode::BOTH;
-    input_message_type input_message     = input_message_type::NONE;
-};
-
-static std::string read_file(const std::string & path) {
-    std::ifstream fin(path, std::ios::binary);
-    if (!fin.is_open()) {
-        throw std::runtime_error("Could not open file: " + path);
-    }
-    std::ostringstream buf;
-    buf << fin.rdbuf();
-    return buf.str();
-}
-
-static std::string read_gguf_chat_template(const std::string & path) {
-    struct gguf_init_params params = { /*no_alloc =*/true,  // We only need metadata, not tensor data
-                                       /*ctx=*/nullptr };
-
-    struct gguf_context * ctx = gguf_init_from_file(path.c_str(), params);
-    if (ctx == nullptr) {
-        throw std::runtime_error("Could not open GGUF file: " + path);
-    }
-
-    const char * key    = "tokenizer.chat_template";
-    int64_t      key_id = gguf_find_key(ctx, key);
-
-    if (key_id == -1) {
-        gguf_free(ctx);
-        throw std::runtime_error("GGUF file does not contain chat template key: " + std::string(key));
-    }
-
-    const char * template_str = gguf_get_val_str(ctx, key_id);
-    if (template_str == nullptr) {
-        gguf_free(ctx);
-        throw std::runtime_error("GGUF file contains chat template key but value is null");
-    }
-
-    std::string result = template_str;
-    gguf_free(ctx);
-    return result;
-}
-
-static void print_usage(const char * program_name) {
-    LOG_ERR("Usage: %s <template_or_gguf_path> [options]\n", program_name);
-    LOG_ERR("\nOptions:\n");
-    LOG_ERR("  --no-tools              Disable tool definitions\n");
-    LOG_ERR("  --force-tool-call       Set tool calls to forced\n");
-    LOG_ERR("  --parallel-tool-calls=0|1 Set parallel_tool_calls (default: 1)\n");
-    LOG_ERR("  --generation-prompt=0|1 Set add_generation_prompt (default: 1)\n");
-    LOG_ERR("  --enable-reasoning=0|1  Enable reasoning parsing (default: 1)\n");
-    LOG_ERR("  --output=MODE           Output mode: analysis, template, both (default: both)\n");
-    LOG_ERR("  --debug-jinja           Enable Jinja fine-grained debug\n");
-    LOG_ERR("  --input-message=TYPE    Message type to render:\n");
-    LOG_ERR("                          content_only, reasoning_content, tool_call_only,\n");
-    LOG_ERR("                          content_tool_call, reasoning_tool_call,\n");
-    LOG_ERR("                          content_fake_tool_call, all\n");
-    LOG_ERR("\nExamples:\n");
-    LOG_ERR("  %s template.jinja --input-message=all --generation-prompt=1\n", program_name);
-    LOG_ERR("  %s template.jinja --output=template --input-message=tool_call_only\n", program_name);
-}
-
-static bool parse_bool_option(const std::string & value) {
-    return value == "1" || value == "true" || value == "yes";
-}
-
-static bool parse_options(int argc, char ** argv, debug_options & opts) {
-    if (argc < 2) {
-        print_usage(argv[0]);
-        return false;
-    }
-
-    opts.template_path = argv[1];
-
-    for (int i = 2; i < argc; ++i) {
-        std::string arg = argv[i];
-
-        if (arg == "--force-tool-call") {
-            opts.force_tool_call = true;
-        } else if (arg == "--debug-jinja") {
-            opts.debug_jinja = true;
-        } else if (arg == "--no-tools") {
-            opts.with_tools = false;
-        } else if (arg.rfind("--parallel-tool-calls=", 0) == 0) {
-            opts.parallel_tool_calls = parse_bool_option(arg.substr(22));
-        } else if (arg.rfind("--generation-prompt=", 0) == 0) {
-            opts.generation_prompt = parse_bool_option(arg.substr(20));
-        } else if (arg.rfind("--enable-reasoning=", 0) == 0) {
-            opts.enable_reasoning = parse_bool_option(arg.substr(19));
-        } else if (arg.rfind("--output=", 0) == 0) {
-            std::string mode = arg.substr(9);
-            if (mode == "analysis") {
-                opts.mode = output_mode::ANALYSIS;
-            } else if (mode == "template") {
-                opts.mode = output_mode::TEMPLATE;
-            } else if (mode == "both") {
-                opts.mode = output_mode::BOTH;
-            } else {
-                LOG_ERR("Unknown output mode: %s\n", mode.c_str());
-                return false;
-            }
-        } else if (arg.rfind("--input-message=", 0) == 0) {
-            std::string type = arg.substr(16);
-            if (type == "content_only") {
-                opts.input_message = input_message_type::CONTENT_ONLY;
-            } else if (type == "reasoning_content") {
-                opts.input_message = input_message_type::REASONING_CONTENT;
-            } else if (type == "tool_call_only") {
-                opts.input_message = input_message_type::TOOL_CALL_ONLY;
-            } else if (type == "content_tool_call") {
-                opts.input_message = input_message_type::CONTENT_TOOL_CALL;
-            } else if (type == "reasoning_tool_call") {
-                opts.input_message = input_message_type::REASONING_TOOL_CALL;
-            } else if (type == "content_fake_tool_call") {
-                opts.input_message = input_message_type::CONTENT_FAKE_TOOL_CALL;
-            } else if (type == "all") {
-                opts.input_message = input_message_type::ALL;
-            } else {
-                LOG_ERR("Unknown input message type: %s\n", type.c_str());
-                return false;
-            }
-        } else {
-            LOG_ERR("Unknown option: %s\n", arg.c_str());
-            print_usage(argv[0]);
-            return false;
-        }
-    }
-
-    return true;
-}
-
-static json build_user_message() {
-    return json{
-        { "role",    "user"                               },
-        { "content", "Hello, please help me with a task." }
-    };
-}
-
-static json build_content_only_message() {
-    return json{
-        { "role",    "assistant"                                   },
-        { "content", "Hello! I'm here to help you with your task." }
-    };
-}
-
-static json build_reasoning_content_message() {
-    return json{
-        { "role",              "assistant"                                                               },
-        { "content",           "Hello! I'm here to help you with your task."                             },
-        { "reasoning_content", "The user is greeting me and asking for help. I should respond politely." }
-    };
-}
-
-static json build_tool_call_only_message() {
-    return json{
-        { "role",       "assistant"      },
-        { "content",    nullptr          },
-        { "tool_calls",
-         json::array({ json{
-              { "type", "function" },
-              { "function", json{ { "name", "test_function_name" },
-                                  { "arguments", json::object({ { "param1", "value1" }, { "param2", "value2" } }) } } },
-              { "id", "123456789" } } }) }
-    };
-}
-
-static json build_content_tool_call_message() {
-    return json{
-        { "role",       "assistant"                                                                              },
-        { "content",    "I'll help you by calling a function."                                                   },
-        { "tool_calls",
-         json::array({ json{
-              { "type", "function" },
-              { "function",
-                json{ { "name", "test_function_name" },
-                      { "arguments", json::object({ { "param1", "value1" }, { "param2", "value2" } }) } } } } }) }
-    };
-}
-
-static json build_reasoning_tool_call_message() {
-    return json{
-        { "role",              "assistant"                                                                       },
-        { "content",           nullptr                                                                           },
-        { "reasoning_content", "I need to call a function to help with this task."                               },
-        { "tool_calls",
-         json::array({ json{
-              { "type", "function" },
-              { "function",
-                json{ { "name", "test_function_name" },
-                      { "arguments", json::object({ { "param1", "value1" }, { "param2", "value2" } }) } } } } }) }
-    };
-}
-
-static json build_content_fake_tool_call_message() {
-    // This message has content but NO tool_calls field
-    // It's used to test if a template renders tool definitions but not tool calls
-    return json{
-        { "role",    "assistant"                            },
-        { "content", "I'll help you by calling a function." }
-    };
-}
-
-static json build_tools_definition() {
-    json parameters_schema                    = json::object();
-    parameters_schema["type"]                 = "object";
-    parameters_schema["properties"]           = json::object();
-    parameters_schema["properties"]["param1"] = json::object({
-        { "type",        "string"          },
-        { "description", "First parameter" }
-    });
-    parameters_schema["properties"]["param2"] = json::object({
-        { "type",        "string"           },
-        { "description", "Second parameter" }
-    });
-    parameters_schema["required"]             = json::array({ "param1" });
-
-    return json::array({
-        json{ { "type", "function" },
-             { "function", json{ { "name", "test_function_name" },
-                                  { "description", "A test function for debugging" },
-                                  { "parameters", parameters_schema } } } }
-    });
-}
-
-static void render_scenario(const common_chat_template & tmpl,
-                            const std::string &          scenario_name,
-                            const json &                 messages,
-                            const json &                 tools,
-                            bool                         add_generation_prompt,
-                            bool                         enable_thinking) {
-    LOG_ERR("\n=== Scenario: %s ===\n", scenario_name.c_str());
-    LOG_ERR("add_generation_prompt: %s, enable_thinking: %s\n", add_generation_prompt ? "true" : "false",
-            enable_thinking ? "true" : "false");
-
-    // When add_generation_prompt is true, add a trailing user message to trigger the prompt
-    json final_messages = messages;
-    if (add_generation_prompt && !messages.empty() && messages.back().value("role", "") == "assistant") {
-        final_messages.push_back(json{
-            { "role",    "user"                                       },
-            { "content", "Now please continue with another response." }
-        });
-    }
-
-    LOG_ERR("Messages:\n%s\n", final_messages.dump(2).c_str());
-
-    try {
-        autoparser::generation_params inputs;
-        inputs.messages                         = final_messages;
-        inputs.add_generation_prompt            = add_generation_prompt;
-        inputs.extra_context["enable_thinking"] = enable_thinking;
-
-        if (!tools.is_null() && tools.is_array() && !tools.empty()) {
-            inputs.tools = tools;
-        }
-
-        std::string output = common_chat_template_direct_apply(tmpl, inputs);
-
-        LOG_ERR("\n--- Rendered Output ---\n");
-        LOG_ERR("%s\n", output.c_str());
-        LOG_ERR("--- End Output (length: %zu) ---\n", output.length());
-    } catch (const std::exception & e) {
-        LOG_ERR("Rendering failed: %s\n", e.what());
-    }
-}
-
-static void render_all_scenarios(const common_chat_template & tmpl,
-                                 const json &                 tools,
-                                 bool                         add_generation_prompt,
-                                 bool                         enable_thinking,
-                                 input_message_type           message_type) {
-    json user_msg = build_user_message();
-
-    auto render_if = [&](input_message_type type, const std::string & name, const json & assistant_msg) {
-        if (message_type == input_message_type::ALL || message_type == type) {
-            json messages = json::array({ user_msg, assistant_msg });
-            render_scenario(tmpl, name, messages, tools, add_generation_prompt, enable_thinking);
-        }
-    };
-
-    render_if(input_message_type::CONTENT_ONLY, "content_only", build_content_only_message());
-    render_if(input_message_type::REASONING_CONTENT, "reasoning_content", build_reasoning_content_message());
-    render_if(input_message_type::TOOL_CALL_ONLY, "tool_call_only", build_tool_call_only_message());
-    render_if(input_message_type::CONTENT_TOOL_CALL, "content_tool_call", build_content_tool_call_message());
-    render_if(input_message_type::REASONING_TOOL_CALL, "reasoning_tool_call", build_reasoning_tool_call_message());
-    render_if(input_message_type::CONTENT_FAKE_TOOL_CALL, "content_fake_tool_call",
-              build_content_fake_tool_call_message());
-
-    // Also render with add_generation_prompt=true to show the prompt ending
-    if (message_type == input_message_type::ALL) {
-        LOG_ERR("\n\n=== Generation Prompt Scenarios (add_generation_prompt=true) ===\n");
-
-        json prompt_messages = json::array({ user_msg });
-        render_scenario(tmpl, "generation_prompt_only", prompt_messages, tools, true, enable_thinking);
-
-        // With enable_thinking toggled
-        render_scenario(tmpl, "generation_prompt_thinking_disabled", prompt_messages, tools, true, false);
-    }
-}
-
-static autoparser::generation_params prepare_params(const debug_options & opts, const json & tools) {
-    autoparser::generation_params params;
-    params.messages         = json::array({ build_user_message() });
-    params.reasoning_format = opts.enable_reasoning ? COMMON_REASONING_FORMAT_DEEPSEEK : COMMON_REASONING_FORMAT_NONE;
-    params.enable_thinking  = opts.enable_reasoning;
-    params.add_generation_prompt = opts.generation_prompt;
-
-    if (opts.with_tools) {
-        params.tools       = tools;
-        params.tool_choice = opts.force_tool_call ? COMMON_CHAT_TOOL_CHOICE_REQUIRED : COMMON_CHAT_TOOL_CHOICE_AUTO;
-    } else {
-        params.tools       = json();
-        params.tool_choice = COMMON_CHAT_TOOL_CHOICE_NONE;
-    }
-    params.parallel_tool_calls = opts.parallel_tool_calls;
-    return params;
-}
-
-int main(int argc, char ** argv) {
-    // Set log level to most verbose to capture all debug output
-    common_log_set_verbosity_thold(99);
-
-    debug_options opts;
-    if (!parse_options(argc, argv, opts)) {
-        return 1;
-    }
-
-    if (opts.debug_jinja || std::getenv("LLAMA_DEBUG_JINJA") != nullptr) {
-        jinja::enable_debug(true);
-    }
-
-    std::string template_source;
-    try {
-        // Check if the file is a GGUF file
-        if (opts.template_path.size() >= 5 &&
-            opts.template_path.compare(opts.template_path.size() - 5, 5, ".gguf") == 0) {
-            template_source = read_gguf_chat_template(opts.template_path);
-        } else {
-            template_source = read_file(opts.template_path);
-        }
-    } catch (const std::exception & e) {
-        LOG_ERR("Error reading template: %s\n", e.what());
-        return 1;
-    }
-
-    LOG_ERR("Analyzing template: %s\n", opts.template_path.c_str());
-    LOG_ERR("Options: with_tools=%s, generation_prompt=%s, enable_reasoning=%s\n", opts.with_tools ? "true" : "false",
-            opts.generation_prompt ? "true" : "false", opts.enable_reasoning ? "true" : "false");
-
-    try {
-        common_chat_template chat_template(template_source, "", "");
-
-        json tools = opts.with_tools ? build_tools_definition() : json();
-
-        autoparser::generation_params params = prepare_params(opts, tools);
-        common_chat_params            parser_data;
-        if (std::optional<common_chat_params> spec_tmpl =
-                common_chat_try_specialized_template(chat_template, template_source, params)) {
-            LOG_ERR("\n");
-            LOG_ERR("This template uses a specialized parser, analysis results will not be available.\n");
-            parser_data = *spec_tmpl;
-        } else {
-            // Render template scenarios if requested
-            if (opts.input_message != input_message_type::NONE &&
-                (opts.mode == output_mode::TEMPLATE || opts.mode == output_mode::BOTH)) {
-                LOG_ERR("\n");
-                LOG_ERR("================================================================================\n");
-                LOG_ERR("                         TEMPLATE RENDERING OUTPUT\n");
-                LOG_ERR("================================================================================\n");
-
-                render_all_scenarios(chat_template, tools, opts.generation_prompt, opts.enable_reasoning,
-                                     opts.input_message);
-            }
-
-            // Output analysis if requested
-            if (opts.mode == output_mode::ANALYSIS || opts.mode == output_mode::BOTH) {
-                LOG_ERR("\n");
-                LOG_ERR("================================================================================\n");
-                LOG_ERR("                           TEMPLATE ANALYSIS\n");
-                LOG_ERR("================================================================================\n");
-
-                autoparser::autoparser analysis;
-                analysis.analyze_template(chat_template);
-
-                // Generate Parser
-                parser_data = autoparser::peg_generator::generate_parser(chat_template, params, analysis);
-            }
-        }
-
-        if (!std::empty(parser_data.parser)) {
-            LOG_ERR("\n=== Generated Parser ===\n");
-            common_peg_arena arena;
-            arena.load(parser_data.parser);
-            LOG_ERR("%s\n", arena.dump(arena.root()).c_str());
-
-            LOG_ERR("\n=== Generated Grammar ===\n");
-            LOG_ERR("%s\n", parser_data.grammar.c_str());
-
-            LOG_ERR("\n=== Generated Lazy Grammar ===\n");
-            LOG_ERR("%d\n", parser_data.grammar_lazy);
-
-            LOG_ERR("\n=== Generated Grammar Triggers ===\n");
-            for (const common_grammar_trigger & cgt : parser_data.grammar_triggers) {
-                LOG_ERR("Token: %d | Type: %d | Value: %s\n", cgt.token, cgt.type, cgt.value.c_str());
-            }
-
-            LOG_ERR("\n=== Preserved Tokens ===\n");
-            for (const std::string & token : parser_data.preserved_tokens) {
-                LOG_ERR("  '%s'\n", token.c_str());
-            }
-
-            if (!parser_data.grammar.empty()) {
-                LOG_ERR("\n=== Verifying created grammar ===\n");
-                auto * grammar = llama_grammar_init_impl(nullptr, parser_data.grammar.c_str(), "root",
-                                                         parser_data.grammar_lazy, nullptr, 0, nullptr, 0);
-                if (grammar != nullptr) {
-                    LOG_ERR("\n=== Grammar successfully created ===\n");
-                }
-            }
-        }
-    } catch (const std::exception & e) {
-        LOG_ERR("Analysis failed: %s\n", e.what());
-        return 1;
-    }
-
-    return 0;
-}
diff --git a/tools/server/server-chat.cpp b/tools/server/server-chat.cpp
index 0322e54cc..a6fe3c6ba 100644
--- a/tools/server/server-chat.cpp
+++ b/tools/server/server-chat.cpp
@@ -153,7 +153,7 @@ json server_chat_convert_responses_to_chatcmpl(const json & response_body) {
                         prev_msg["content"] = json::array();
                     }
                     auto & prev_content = prev_msg["content"];
-                    prev_content.insert(prev_content.end(), chatcmpl_content.begin(), chatcmpl_content.end());
+                    prev_content.insert(chatcmpl_content);
                 } else {
                     item.erase("status");
                     item.erase("type");
diff --git a/tools/server/server-chat.h b/tools/server/server-chat.h
index 102eae688..86b842650 100644
--- a/tools/server/server-chat.h
+++ b/tools/server/server-chat.h
@@ -6,9 +6,7 @@
 #include "server-common.h"
 #include "server-http.h"
 
-#include <nlohmann/json_fwd.hpp>
-
-using json = nlohmann::ordered_json;
+#include "json.h"
 
 // Convert OpenAI Responses API format to OpenAI Chat Completions API format
 json server_chat_convert_responses_to_chatcmpl(const json & body);
diff --git a/tools/server/server-common.cpp b/tools/server/server-common.cpp
index 585f65e83..4f5b8202a 100644
--- a/tools/server/server-common.cpp
+++ b/tools/server/server-common.cpp
@@ -1540,7 +1540,7 @@ std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int i
 }
 
 std::string safe_json_to_str(const json & data) {
-    return data.dump(-1, ' ', false, json::error_handler_t::replace);
+    return data.dump_safe();
 }
 
 // TODO: reuse llama_detokenize
diff --git a/tools/server/server-common.h b/tools/server/server-common.h
index 6488be344..f8ea82ef4 100644
--- a/tools/server/server-common.h
+++ b/tools/server/server-common.h
@@ -6,8 +6,7 @@
 #include "chat.h"
 #include "mtmd.h"
 
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <atomic>
 #include <chrono>
@@ -19,7 +18,7 @@
 #include <string>
 #include <vector>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 #define SLT_DBG(slot, fmt, ...) LOG_DBG("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, ((slot).task ? (slot).task->id : -1), __VA_ARGS__)
 #define SLT_TRC(slot, fmt, ...) LOG_TRC("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, ((slot).task ? (slot).task->id : -1), __VA_ARGS__)
@@ -42,9 +41,9 @@ static T json_value(const json & body, const std::string & key, const T & defaul
     // Fallback null to default value
     if (body.contains(key) && !body.at(key).is_null()) {
         try {
-            return body.at(key);
-        } catch (NLOHMANN_JSON_NAMESPACE::detail::type_error const & err) {
-            LOG_WRN("Wrong type supplied for parameter '%s'. Expected '%s', using default value: %s\n", key.c_str(), json(default_value).type_name(), err.what());
+            return body.at(key).get<T>();
+        } catch (const common_json_error & err) {
+            LOG_WRN("Wrong type supplied for parameter '%s', using default value: %s\n", key.c_str(), err.what());
             return default_value;
         }
     } else {
diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
index 36d982832..a9edbd7be 100644
--- a/tools/server/server-context.cpp
+++ b/tools/server/server-context.cpp
@@ -35,8 +35,6 @@
 #include <windows.h>
 #endif
 
-using json = nlohmann::ordered_json;
-
 constexpr int HTTP_POLLING_SECONDS = 1;
 
 static common_speculative_output_limits server_output_limits(const common_params & params) {
@@ -657,14 +655,14 @@ struct server_slot {
             res["n_prompt_tokens_processed"] = stats.n_prompt_processed;
             res["n_prompt_tokens_cache"]     = stats.n_prompt_cached;
             res["params"] = ptask->params.to_json(only_metrics);
-            res["next_token"] = {
+            res["next_token"] = json::array({
                 {
                     {"has_next_token", has_next_token},
                     {"has_new_line",   has_new_line},
                     {"n_remain",       n_remaining()},
                     {"n_decoded",      stats.n_gen},
                 }
-            };
+            });
 
             if (!only_metrics) {
                 res["prompt"] = ptask->tokens.detokenize(ctx_tgt, true);
@@ -860,8 +858,10 @@ private:
     // slots / clients
     std::vector<server_slot> slots;
 
-    int trace = 0;
-    int slots_debug = 0;
+    int trace = 0;        // env: LLAMA_TRACE
+    int slots_debug = 0;  // env: LLAMA_SERVER_SLOTS_DEBUG
+    int slots_n_diff = 0; // env: LLAMA_SERVER_SLOTS_N_DIFF
+
     int n_empty_consecutive = 0;
 
     std::unique_ptr<server_prompt_cache> prompt_cache;
@@ -1249,6 +1249,15 @@ private:
             }
         }
 
+        {
+            const char * LLAMA_SERVER_SLOTS_N_DIFF = getenv("LLAMA_SERVER_SLOTS_N_DIFF");
+            slots_n_diff = LLAMA_SERVER_SLOTS_N_DIFF ? atoi(LLAMA_SERVER_SLOTS_N_DIFF) : 0;
+
+            if (slots_n_diff) {
+                SRV_WRN("LLAMA_SERVER_SLOTS_N_DIFF = %d\n", slots_n_diff);
+            }
+        }
+
         // the update_slots() logic will always submit a maximum of n_batch or n_parallel tokens
         // note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used)
         {
@@ -3181,8 +3190,8 @@ private:
                                 // when the prompt prefix does not match, print the tokens around the mismatch
                                 // this is useful for debugging prompt caching
                                 if (slots_debug) {
-                                    const int np0 = std::max<int>(n_past - 4, 0);
-                                    const int np1 = std::min<int>(n_past + 6, std::min(slot.prompt.tokens.size(), slot.task->tokens.size()));
+                                    const int np0 = std::max<int>(n_past - slots_n_diff, 0);
+                                    const int np1 = std::min<int>(n_past + slots_n_diff + 2, std::min(slot.prompt.tokens.size(), slot.task->tokens.size()));
 
                                     std::stringstream ss0;
                                     std::stringstream ss1;
@@ -4165,7 +4174,8 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
         // tasks.reserve(inputs.size()); // TODO: this is inaccurate due to child tasks
 
         // message delimiters for checkpointing
-        auto delimiters = common_chat_msg_delimiters_parse(json_value(data, "message_delimiters", json::array()));
+        json delims = json_value(data, "message_delimiters", json::array());
+        auto delimiters = common_chat_msg_delimiters_parse(delims);
         delimiters.tokenize(ctx_server.vocab);
 
         for (size_t i = 0; i < inputs.size(); i++) {
@@ -4428,8 +4438,8 @@ static json get_res_model_info(const server_context_meta & meta) {
 static json get_res_models(const server_context_meta & meta) {
     // note: do NOT use ctx_server here, otherwise it's not possible to use this during sleep
 
-    return {
-        {"models", {
+    return json{
+        {"models", json::array({
             {
                 {"name",  meta.model_name},
                 {"model", meta.model_name},
@@ -4438,23 +4448,23 @@ static json get_res_models(const server_context_meta & meta) {
                 {"digest", ""}, // dummy value, llama.cpp does not support managing model file's hash
                 {"type", "model"},
                 {"description", ""},
-                {"tags", {""}},
-                {"capabilities", meta.has_mtmd ? json({"completion","multimodal"}) : json({"completion"})},
+                {"tags", json::array({""})},
+                {"capabilities", meta.has_mtmd ? json::array({"completion","multimodal"}) : json::array({"completion"})},
                 {"parameters", ""},
                 {"details", {
                     {"parent_model", ""},
                     {"format", "gguf"},
                     {"family", ""},
-                    {"families", {""}},
+                    {"families", json::array({""})},
                     {"parameter_size", ""},
                     {"quantization_level", ""}
                 }}
             }
-        }},
+        })},
         {"object", "list"},
-        {"data", {
+        {"data", json::array({
             get_res_model_info(meta),
-        }}
+        })}
     };
 }
 
@@ -4990,7 +5000,7 @@ void server_routes::init_routes() {
 
         std::string content;
         if (body.count("tokens") != 0) {
-            const llama_tokens tokens = body.at("tokens");
+            const llama_tokens tokens = body.at("tokens").get<llama_tokens>();
             content = tokens_to_str(ctx_server.vocab, tokens);
         }
 
@@ -5297,7 +5307,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_embeddings_impl(cons
 
     int embd_normalize = params.embd_normalize;
     if (body.count("embd_normalize") != 0) {
-        embd_normalize = body.at("embd_normalize");
+        embd_normalize = body.at("embd_normalize").get<int>();
         if (meta->pooling_type == LLAMA_POOLING_TYPE_NONE) {
             SRV_DBG("embd_normalize is not supported by pooling type %d, ignoring it\n", meta->pooling_type);
         }
diff --git a/tools/server/server-context.h b/tools/server/server-context.h
index 764df0e08..5d464b8e8 100644
--- a/tools/server/server-context.h
+++ b/tools/server/server-context.h
@@ -4,7 +4,7 @@
 #include "server-task.h"
 #include "server-queue.h"
 
-#include <nlohmann/json_fwd.hpp>
+#include "json.h"
 
 #include <cstddef>
 #include <memory>
diff --git a/tools/server/server-models.cpp b/tools/server/server-models.cpp
index d60545194..db0fac995 100644
--- a/tools/server/server-models.cpp
+++ b/tools/server/server-models.cpp
@@ -2462,7 +2462,7 @@ server_http_proxy::server_http_proxy(
     bool has_files = !files.empty();
 
     if (has_files) {
-        json form_fields = json::parse(body, nullptr, false);
+        json form_fields = json::parse_no_throw(body);
         if (!form_fields.is_discarded()) {
             auto boundary = generate_multipart_boundary();
             effective_body = build_multipart_body(form_fields, files, boundary);
diff --git a/tools/server/server-schema.cpp b/tools/server/server-schema.cpp
index 5d7fa6ae6..64b925129 100644
--- a/tools/server/server-schema.cpp
+++ b/tools/server/server-schema.cpp
@@ -503,7 +503,7 @@ std::vector<std::unique_ptr<field>> make_llama_cmpl_schema(const common_params &
         ->set_handler([&](field_eval_context & ctx, const json & data) {
             const auto & samplers = data.at("samplers");
             if (samplers.is_array()) {
-                ctx.params.sampling.samplers = common_sampler_types_from_names(samplers);
+                ctx.params.sampling.samplers = common_sampler_types_from_names(samplers.get<std::vector<std::string>>());
             } else if (samplers.is_string()) {
                 ctx.params.sampling.samplers = common_sampler_types_from_chars(samplers.get<std::string>());
             }
@@ -580,8 +580,7 @@ static void handle_with_catch(const char * name, std::function<void()> func) {
 
 // treat a null value as absent so clients can send null to request the server default
 static bool has_value(const json & data, const char * n) {
-    auto it = data.find(n);
-    return it != data.end() && !it->is_null();
+    return data.contains(n) && !data.at(n).is_null();
 }
 
 template <typename T>
diff --git a/tools/server/server-task.cpp b/tools/server/server-task.cpp
index 258cdcf8f..0d3beb313 100644
--- a/tools/server/server-task.cpp
+++ b/tools/server/server-task.cpp
@@ -12,8 +12,6 @@
 
 #include <sstream>
 
-using json = nlohmann::ordered_json;
-
 //
 // task_params
 //
@@ -304,7 +302,7 @@ json completion_token_output::probs_vector_to_json(const std::vector<completion_
 }
 
 float completion_token_output::logarithm(float x) {
-    // nlohmann::json converts -inf to null, so we need to prevent that
+    // the JSON library converts -inf to null, so we need to prevent that
     return x == 0.0f ? std::numeric_limits<float>::lowest() : std::log(x);
 }
 
@@ -407,7 +405,7 @@ json server_task_result_cmpl_final::to_json_oaicompat() {
         res["__verbose"] = to_json_non_oaicompat();
     }
     if (stats.is_set()) {
-        res.push_back({"timings", stats.to_json()});
+        res["timings"] = stats.to_json();
     }
 
     return res;
@@ -455,7 +453,7 @@ json server_task_result_cmpl_final::to_json_oaicompat_chat() {
         res["__verbose"] = to_json_non_oaicompat();
     }
     if (stats.is_set()) {
-        res.push_back({"timings", stats.to_json()});
+        res["timings"] = stats.to_json();
     }
 
     return res;
@@ -516,7 +514,7 @@ json server_task_result_cmpl_final::to_json_oaicompat_chat_stream() {
     }
 
     if (stats.is_set()) {
-        deltas.back().push_back({"timings", stats.to_json()});
+        deltas.back()["timings"] = stats.to_json();
     }
 
     // extra fields for debugging purposes
@@ -709,7 +707,7 @@ json server_task_result_cmpl_final::to_json_oaicompat_resp_stream() {
     });
 
     if (stats.is_set()) {
-        server_sent_events.back().at("data").push_back({"timings", stats.to_json()});
+        server_sent_events.back().at("data")["timings"] = stats.to_json();
     }
 
     return server_sent_events;
@@ -1061,10 +1059,10 @@ json server_task_result_cmpl_partial::to_json_non_oaicompat() {
     };
     // populate the timings object when needed (usually for the last response or with timings_per_token enabled)
     if (stats.is_set()) {
-        res.push_back({"timings", stats.to_json()});
+        res["timings"] = stats.to_json();
     }
     if (is_progress) {
-        res.push_back({"prompt_progress", progress.to_json()});
+        res["prompt_progress"] = progress.to_json();
     }
     if (!prob_output.probs.empty()) {
         res["completion_probabilities"] = completion_token_output::probs_vector_to_json({prob_output}, post_sampling_probs);
@@ -1101,10 +1099,10 @@ json server_task_result_cmpl_partial::to_json_oaicompat() {
         res["__verbose"] = to_json_non_oaicompat();
     }
     if (stats.is_set()) {
-        res.push_back({"timings", stats.to_json()});
+        res["timings"] = stats.to_json();
     }
     if (is_progress) {
-        res.push_back({"prompt_progress", progress.to_json()});
+        res["prompt_progress"] = progress.to_json();
     }
 
     return res;
@@ -1155,10 +1153,10 @@ json server_task_result_cmpl_partial::to_json_oaicompat_chat() {
         }
 
         if (stats.is_set()) {
-            last_json.push_back({"timings", stats.to_json()});
+            last_json["timings"] = stats.to_json();
         }
         if (is_progress) {
-            last_json.push_back({"prompt_progress", progress.to_json()});
+            last_json["prompt_progress"] = progress.to_json();
         }
     }
 
@@ -1305,10 +1303,10 @@ json server_task_result_cmpl_partial::to_json_oaicompat_resp() {
     if (!events.empty()) {
         json & data = events.back().at("data");
         if (stats.is_set()) {
-            data.push_back({"timings", stats.to_json()});
+            data["timings"] = stats.to_json();
         }
         if (is_progress) {
-            data.push_back({"prompt_progress", progress.to_json()});
+            data["prompt_progress"] = progress.to_json();
         }
     }
 
diff --git a/tools/server/server-task.h b/tools/server/server-task.h
index 25ff01512..9c99143f8 100644
--- a/tools/server/server-task.h
+++ b/tools/server/server-task.h
@@ -11,7 +11,6 @@
 // TODO: prevent including the whole server-common.h as we only use server_tokens
 #include "server-common.h"
 
-using json = nlohmann::ordered_json;
 
 enum server_task_type {
     SERVER_TASK_TYPE_COMPLETION,
diff --git a/tools/server/server-tools.cpp b/tools/server/server-tools.cpp
index b5c5c078a..12e9dbb8c 100644
--- a/tools/server/server-tools.cpp
+++ b/tools/server/server-tools.cpp
@@ -2156,7 +2156,7 @@ void server_tools::setup(const std::vector<std::string> & enabled_tools,
                 res->status = 200;
                 res->data   = safe_json_to_str(result);
             }
-        } catch (const json::exception & e) {
+        } catch (const common_json_error & e) {
             res->status = 400;
             res->data   = safe_json_to_str(format_error_response(e.what(), ERROR_TYPE_INVALID_REQUEST));
         } catch (const std::invalid_argument & e) {
diff --git a/tools/server/tests/unit/test_slot_save.py b/tools/server/tests/unit/test_slot_save.py
index 05acb1be1..5af61d70d 100644
--- a/tools/server/tests/unit/test_slot_save.py
+++ b/tools/server/tests/unit/test_slot_save.py
@@ -319,7 +319,6 @@ def test_slot_save_restore_with_two_images(mmproj_server):
         "prompt": prompt,
     })
     assert res.status_code == 200
-    content = res.body["content"]
     prompt_n_full = res.body["timings"]["prompt_n"]
     assert prompt_n_full > 64
 
@@ -345,6 +344,26 @@ def test_slot_save_restore_with_two_images(mmproj_server):
     assert res.status_code == 200
     assert res.body["timings"]["cache_n"] == prompt_n_full - 1
     assert res.body["timings"]["prompt_n"] == 1
+    content = res.body["content"]
+
+    res = server.make_request("POST", "/slots/1?action=restore", data={
+        "filename": "mm_slot_two_images.bin",
+    })
+    assert res.status_code == 200
+    assert res.body["n_restored"] == n_saved
+
+    res = server.make_request("POST", "/completions", data={
+        "temperature": 0.0,
+        "top_k": 1,
+        "id_slot": 0,
+        "cache_prompt": True,
+        "prompt": prompt,
+    })
+    assert res.status_code == 200
+    assert res.body["timings"]["cache_n"] == prompt_n_full - 1
+    assert res.body["timings"]["prompt_n"] == 1
+    content = res.body["content"]
+
     assert res.body["content"] == content
 
 
diff --git a/tools/server/tests/unit/test_vision_api.py b/tools/server/tests/unit/test_vision_api.py
index d74cc3a43..8b01c5372 100644
--- a/tools/server/tests/unit/test_vision_api.py
+++ b/tools/server/tests/unit/test_vision_api.py
@@ -121,7 +121,7 @@ def test_vision_chat_completion_token_count():
     "prompt, image_data, success, re_content",
     [
         # test model is trained on CIFAR-10, but it's quite dumb due to small size
-        ("What is this: <__media__>\n", "IMG_BASE64_0",         True, "(cat)+"),
+        ("What is this: <__media__>\n", "IMG_BASE64_0",         True, "(cat)+|(automobile)+"),
         ("What is this: <__media__>\n", "IMG_BASE64_1",         True, "(frog)+"),
         ("What is this: <__media__>\n", "malformed",            False, None), # non-image data
         ("What is this:\n",             "",                     False, None), # empty string
diff --git a/tools/server/tests/utils.py b/tools/server/tests/utils.py
index 9171dbc02..a0d2dfa3c 100644
--- a/tools/server/tests/utils.py
+++ b/tools/server/tests/utils.py
@@ -623,7 +623,7 @@ class ServerPreset:
         server.model_hf_repo = "ggml-org/tinygemma3-GGUF:Q8_0"
         server.model_alias = "tinygemma3"
         server.n_ctx = 1024
-        server.n_batch = 32
+        server.n_batch = 512
         server.n_slots = 2
         server.n_predict = 4
         server.seed = 42
diff --git a/tools/ui/src/lib/components/app/chat/ChatAttachments/ChatAttachmentsList/ChatAttachmentsList.svelte b/tools/ui/src/lib/components/app/chat/ChatAttachments/ChatAttachmentsList/ChatAttachmentsList.svelte
index 36895c8e7..2de9460aa 100644
--- a/tools/ui/src/lib/components/app/chat/ChatAttachments/ChatAttachmentsList/ChatAttachmentsList.svelte
+++ b/tools/ui/src/lib/components/app/chat/ChatAttachments/ChatAttachmentsList/ChatAttachmentsList.svelte
@@ -3,8 +3,9 @@
 		ChatAttachmentsListItem,
 		DialogChatAttachmentsPreview,
 		DialogMcpResourcePreview,
-		HorizontalScrollCarousel
+		ScrollCarousel
 	} from '$lib/components/app';
+	import { ScrollCarouselVariant } from '$lib/enums';
 	import type { DatabaseMessageExtraMcpResource } from '$lib/types';
 	import { getAttachmentDisplayItems, isMcpPrompt, isMcpResource } from '$lib/utils';
 
@@ -42,7 +43,7 @@
 		uploadedFiles = $bindable([])
 	}: Props = $props();
 
-	let carouselRef: HorizontalScrollCarousel | undefined = $state();
+	let carouselRef: ScrollCarousel | undefined = $state();
 	let mcpResourcePreviewOpen = $state(false);
 	let mcpResourcePreviewExtra = $state<DatabaseMessageExtraMcpResource | null>(null);
 	let previewFocusIndex = $state(0);
@@ -91,11 +92,11 @@
 {#if displayItems.length > 0}
 	<div class={className} {style}>
 		{#if limitToSingleRow}
-			<HorizontalScrollCarousel bind:this={carouselRef}>
+			<ScrollCarousel bind:this={carouselRef} variant={ScrollCarouselVariant.CENTER}>
 				{#each displayItems as item (item.id)}
 					{@render attachmentitem(item)}
 				{/each}
-			</HorizontalScrollCarousel>
+			</ScrollCarousel>
 		{:else}
 			<div class="flex flex-wrap items-start justify-end gap-3">
 				{#each displayItems as item (item.id)}
diff --git a/tools/ui/src/lib/components/app/chat/ChatAttachments/ChatAttachmentsPreview/ChatAttachmentsPreviewThumbnailStrip.svelte b/tools/ui/src/lib/components/app/chat/ChatAttachments/ChatAttachmentsPreview/ChatAttachmentsPreviewThumbnailStrip.svelte
index 366c8372b..f0ea9675f 100644
--- a/tools/ui/src/lib/components/app/chat/ChatAttachments/ChatAttachmentsPreview/ChatAttachmentsPreviewThumbnailStrip.svelte
+++ b/tools/ui/src/lib/components/app/chat/ChatAttachments/ChatAttachmentsPreview/ChatAttachmentsPreviewThumbnailStrip.svelte
@@ -1,7 +1,8 @@
 <script lang="ts">
 	import { FileText, Music, Video } from '@lucide/svelte';
-	import { HorizontalScrollCarousel } from '$lib/components/app/misc';
+	import { ScrollCarousel } from '$lib/components/app';
 	import { ICON_CLASS_DEFAULT, UI_DATA_ATTRS } from '$lib/constants';
+	import { ScrollCarouselVariant } from '$lib/enums';
 
 	interface PreviewItem {
 		id: string;
@@ -33,7 +34,7 @@
 
 {#if items.length > 1}
 	<div class="sticky bottom-0 z-10 mt-4 flex-shrink-0">
-		<HorizontalScrollCarousel class="max-w-full">
+		<ScrollCarousel class="max-w-full" variant={ScrollCarouselVariant.CENTER}>
 			{#each items as item, index (item.id)}
 				<button
 					{...{ [UI_DATA_ATTRS.THUMBNAIL_INDEX]: index }}
@@ -64,6 +65,6 @@
 					{/if}
 				</button>
 			{/each}
-		</HorizontalScrollCarousel>
+		</ScrollCarousel>
 	</div>
 {/if}
diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormInput/ChatFormInputRich.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormInput/ChatFormInputRich.svelte
index 5ce788d5a..629001bb4 100644
--- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormInput/ChatFormInputRich.svelte
+++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormInput/ChatFormInputRich.svelte
@@ -625,7 +625,7 @@
 		}
 
 		if (rootElement && (event.key === 'ArrowLeft' || event.key === 'ArrowRight')) {
-			const isWordJump = (event.altKey || event.ctrlKey) && !event.metaKey;
+			const isWordJump = (event.altKey || event.ctrlKey) && !event.metaKey && !event.shiftKey;
 			const isPlainLeft =
 				event.key === 'ArrowLeft' && !event.altKey && !event.ctrlKey && !event.metaKey;
 
diff --git a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormMcpResourcesList.svelte b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormMcpResourcesList.svelte
index 452fd7a68..18d86bcab 100644
--- a/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormMcpResourcesList.svelte
+++ b/tools/ui/src/lib/components/app/chat/ChatForm/ChatFormMcpResourcesList.svelte
@@ -1,8 +1,6 @@
 <script lang="ts">
-	import {
-		ChatAttachmentsListItemMcpResource,
-		HorizontalScrollCarousel
-	} from '$lib/components/app';
+	import { ChatAttachmentsListItemMcpResource, ScrollCarousel } from '$lib/components/app';
+	import { ScrollCarouselVariant } from '$lib/enums';
 	import { mcpStore } from '$lib/stores';
 
 	interface Props {
@@ -26,7 +24,7 @@
 
 {#if hasAttachments}
 	<div class={className}>
-		<HorizontalScrollCarousel gapSize="2">
+		<ScrollCarousel gapSize="2" variant={ScrollCarouselVariant.CENTER}>
 			{#each attachments as attachment, i (attachment.id)}
 				<ChatAttachmentsListItemMcpResource
 					class={i === 0 ? 'ml-3' : ''}
@@ -35,6 +33,6 @@
 					onclick={() => handleResourceClick(attachment.resource.uri)}
 				/>
 			{/each}
-		</HorizontalScrollCarousel>
+		</ScrollCarousel>
 	</div>
 {/if}
diff --git a/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageAssistant/ChatMessageAssistant.svelte b/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageAssistant/ChatMessageAssistant.svelte
index c92af719a..f0be6e95a 100644
--- a/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageAssistant/ChatMessageAssistant.svelte
+++ b/tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageAssistant/ChatMessageAssistant.svelte
@@ -196,7 +196,7 @@
 		--assistant-min-height-offset: calc(
 			var(--last-user-message-height, 19rem) + var(--chat-form-height, 6rem) +
 				var(--chat-form-bottom-position, 0.5rem) + var(--chat-form-padding-top, 6rem) +
-				var(--assistant-margin-top, 3rem)
+				var(--assistant-margin-top, 3rem) + var(--chat-tabs-offset, 0px)
 		);
 		min-height: calc(100dvh - var(--assistant-min-height-offset));
 
@@ -204,7 +204,7 @@
 			--assistant-min-height-offset: calc(
 				var(--last-user-message-height, 18rem) + var(--chat-form-height, 6rem) +
 					var(--chat-form-bottom-position, 1rem) + var(--chat-form-padding-top, 6rem) +
-					var(--assistant-margin-top, 3rem)
+					var(--assistant-margin-top, 3rem) + var(--chat-tabs-offset, 0px)
 			);
 		}
 	}
diff --git a/tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte b/tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte
index d78de1f85..7afb4c0cf 100644
--- a/tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte
+++ b/tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte
@@ -41,10 +41,7 @@
 	let showDeleteDialog = $state(false);
 	let showEmptyFileDialog = $state(false);
 	let isEmpty = $derived(
-		showCenteredEmpty &&
-			!conversationsStore.activeConversation &&
-			conversationsStore.activeMessages.length === 0 &&
-			!chatStore.isLoading
+		showCenteredEmpty && conversationsStore.activeMessages.length === 0 && !chatStore.isLoading
 	);
 	let activeErrorDialog = $derived(chatStore.errorDialogState);
 	let isServerLoading = $derived(serverStore.loading);
@@ -297,7 +294,7 @@
 	<ServerLoadingSplash />
 {:else}
 	<div
-		class="chat-screen flex grow flex-col min-h-[calc(100dvh-1rem)] md:min-h-full px-4 md:py-0 pt-12 pb-48 md:pb-4"
+		class="chat-screen flex grow flex-col min-h-[calc(100dvh-1rem)] md:min-h-[calc(100dvh-1rem-var(--chat-tabs-offset,0px))] px-4 md:py-0 pt-12 pb-48 md:pb-4"
 		style:--chat-form-bottom-position={chatFormBottomPosition}
 		ondragenter={dragAndDrop.dragHandlers.dragenter}
 		ondragleave={dragAndDrop.dragHandlers.dragleave}
diff --git a/tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenActionScrollDown.svelte b/tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenActionScrollDown.svelte
index 6fd27212d..9d7a6ee55 100644
--- a/tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenActionScrollDown.svelte
+++ b/tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenActionScrollDown.svelte
@@ -14,6 +14,6 @@
 		tooltip="Scroll to bottom"
 		size="lg"
 		iconSize={ICON_CLASS_DEFAULT}
-		class="h-9 w-9 rounded-full bg-accent text-accent-foreground absolute bottom-4 shadow-md"
+		class="h-9 w-9 rounded-full bg-muted/60 border border-border/20 shadow-sm text-accent-foreground absolute bottom-4"
 	/>
 </div>
diff --git a/tools/ui/src/lib/components/app/chat/ChatTabs/ChatTabs.svelte b/tools/ui/src/lib/components/app/chat/ChatTabs/ChatTabs.svelte
new file mode 100644
index 000000000..763241c56
--- /dev/null
+++ b/tools/ui/src/lib/components/app/chat/ChatTabs/ChatTabs.svelte
@@ -0,0 +1,136 @@
+<script lang="ts">
+	import ChatTabsItem from './ChatTabsItem.svelte';
+	import ChatTabsNewChatButton from './ChatTabsNewChatButton.svelte';
+	import { page } from '$app/state';
+	import { ScrollCarousel } from '$lib/components/app';
+	import {
+		CHAT_TABS_MAX_WIDTH,
+		NEW_CHAT_LABEL,
+		NEW_CHAT_TAB_ID,
+		UI_DATA_ATTRS,
+		UNNAMED_CHAT_LABEL
+	} from '$lib/constants';
+	import { useScrollCarousel } from '$lib/hooks/use-scroll-carousel.svelte';
+	import { chatStore, conversationsStore, tabsStore, uiStore } from '$lib/stores';
+	import { tick } from 'svelte';
+
+	const carousel = useScrollCarousel();
+
+	let activeId = $derived(page.params.id ?? NEW_CHAT_TAB_ID);
+
+	let tabs = $derived(
+		tabsStore.openTabs.map((id) => ({
+			id,
+			isNewChat: id === NEW_CHAT_TAB_ID,
+			name:
+				id === NEW_CHAT_TAB_ID
+					? NEW_CHAT_LABEL
+					: (conversationsStore.conversations.find((c) => c.id === id)?.name ?? UNNAMED_CHAT_LABEL)
+		}))
+	);
+
+	// hide the New chat button while a new-chat tab is already open
+	let showNewChatButton = $derived(!tabsStore.openTabs.includes(NEW_CHAT_TAB_ID));
+
+	let loadingIds = $derived(new Set(chatStore.getAllLoadingChats()));
+
+	function handleClose(id: string) {
+		void tabsStore.close(id, activeId);
+	}
+
+	function handleStop(id: string, event: MouseEvent) {
+		event.stopPropagation();
+		void chatStore.stopGenerationForChat(id);
+	}
+
+	function handleAuxClick(id: string, event: MouseEvent) {
+		// middle-click closes, like browser tabs
+		if (event.button === 1) {
+			event.preventDefault();
+			handleClose(id);
+		}
+	}
+
+	let previousTabIds = new Set<string>();
+	let previousActiveId: string | null = null;
+
+	$effect(() => {
+		const currentIds = new Set(tabs.map((t) => t.id));
+		const hasAddedTab = tabs.some((t) => !previousTabIds.has(t.id));
+
+		previousTabIds = currentIds;
+
+		const activeChanged = activeId !== previousActiveId;
+
+		previousActiveId = activeId;
+
+		// scroll when the active tab changes (a click) or when a new tab is added
+		if (!hasAddedTab && !activeChanged) return;
+
+		// wait for the new tab to be laid out before scrolling to it
+		void tick().then(() => {
+			const el = carousel.scrollContainer?.querySelector<HTMLElement>(
+				`[${UI_DATA_ATTRS.ACTIVE_TAB}]`
+			);
+
+			if (el) {
+				carousel.scrollToCenter(el);
+			}
+		});
+	});
+</script>
+
+<nav
+	class="group sticky pl-1 top-0 z-10 hidden md:block chat-tabs-fade transition-[padding] duration-200 ease-in-out pt-3.25 {uiStore.isSidebarExpanded
+		? CHAT_TABS_MAX_WIDTH.EXPANDED_SIDEBAR
+		: CHAT_TABS_MAX_WIDTH.COLLAPSED_SIDEBAR}"
+	aria-label="Open conversations"
+>
+	<div class="relative">
+		<ScrollCarousel
+			class="h-10"
+			containerClass="flex h-10 min-w-0 items-center"
+			innerClass="items-center gap-1.25"
+			{carousel}
+		>
+			{#each tabs as tab (tab.id)}
+				<ChatTabsItem
+					{tab}
+					isActive={tab.id === activeId}
+					isLoading={loadingIds.has(tab.id)}
+					onActivate={(id) => tabsStore.activate(id)}
+					onClose={handleClose}
+					onStop={handleStop}
+					onAuxClick={handleAuxClick}
+				/>
+			{/each}
+
+			{#if showNewChatButton}
+				<ChatTabsNewChatButton onclick={() => void conversationsStore.openNewChat()} />
+			{/if}
+		</ScrollCarousel>
+
+		<div
+			class="pointer-events-none absolute inset-y-0 left-0 z-[5] w-8 bg-gradient-to-r from-background to-transparent transition-opacity {carousel.canScrollLeft
+				? 'opacity-100'
+				: 'opacity-0'}"
+		></div>
+		<div
+			class="pointer-events-none absolute inset-y-0 right-0 z-[5] w-8 bg-gradient-to-l from-background to-transparent transition-opacity {carousel.canScrollRight
+				? 'opacity-100'
+				: 'opacity-0'}"
+		></div>
+	</div>
+</nav>
+
+<style>
+	.chat-tabs-fade {
+		background: linear-gradient(
+			to bottom,
+			color-mix(in srgb, var(--background) 100%, transparent) 25%,
+			color-mix(in srgb, var(--background) 80%, transparent) 50%,
+			color-mix(in srgb, var(--background) 40%, transparent) 75%,
+			transparent 100%
+		);
+	}
+</style>
diff --git a/tools/ui/src/lib/components/app/chat/ChatTabs/ChatTabsItem.svelte b/tools/ui/src/lib/components/app/chat/ChatTabs/ChatTabsItem.svelte
new file mode 100644
index 000000000..b76a88efe
--- /dev/null
+++ b/tools/ui/src/lib/components/app/chat/ChatTabs/ChatTabsItem.svelte
@@ -0,0 +1,156 @@
+<script lang="ts">
+	import { Loader2, Square, SquarePen, X } from '@lucide/svelte';
+	import * as Tooltip from '$lib/components/ui/tooltip';
+	import { cn } from '$lib/components/ui/utils';
+	import { ICON_CLASS_SM, ICON_CLASS_XS, ROUTES, UI_DATA_ATTRS } from '$lib/constants';
+	import { RouterService } from '$lib/services/router.service';
+
+	interface Tab {
+		id: string;
+		isNewChat: boolean;
+		name: string;
+	}
+
+	interface Props {
+		tab: Tab;
+		isActive?: boolean;
+		isLoading?: boolean;
+		onActivate?: (id: string) => void;
+		onClose?: (id: string) => void;
+		onStop?: (id: string, event: MouseEvent) => void;
+		onAuxClick?: (id: string, event: MouseEvent) => void;
+	}
+
+	let {
+		isActive = false,
+		isLoading = false,
+		onActivate,
+		onAuxClick,
+		onClose,
+		onStop,
+		tab
+	}: Props = $props();
+
+	let contentOpacity = $derived(isActive ? '' : 'opacity-45 group-hover:opacity-75');
+
+	let href = $derived(tab.isNewChat ? ROUTES.START : RouterService.chat(tab.id));
+
+	function handleActivate(event: MouseEvent) {
+		// let cmd/ctrl/middle-click fall through so the browser keeps its own
+		// behavior (open in a new window); route the plain click ourselves so the
+		// new-chat sentinel and history behave exactly like programmatic nav
+		if (event.metaKey || event.ctrlKey || event.button === 1) return;
+
+		event.preventDefault();
+		onActivate?.(tab.id);
+	}
+
+	// stop/close sit on top of the tab link; swallow their clicks so they do
+	// not also navigate
+	function handleActionClick(event: MouseEvent, action: () => void) {
+		event.preventDefault();
+		event.stopPropagation();
+		action();
+	}
+</script>
+
+<!-- the tab link covers the whole item; stop/close sit on top as siblings so
+     interactive elements are never nested inside the anchor -->
+<div
+	{...{ [UI_DATA_ATTRS.ACTIVE_TAB]: isActive ? 'true' : undefined }}
+	class={cn(
+		'relative flex h-8 max-w-52 min-w-0 shrink-0 items-center gap-1 rounded-lg pr-1 text-sm whitespace-nowrap border backdrop-blur-xl first:ml-2',
+		isLoading ? 'pl-1' : 'pl-3',
+		isActive
+			? 'bg-muted/60 border-border/10 shadow-sm text-accent-foreground hover:bg-primary/15'
+			: 'border-transparent hover:bg-primary/10 hover:border-border/10 hover:shadow-sm'
+	)}
+>
+	<a
+		{href}
+		class="absolute inset-0 z-0 rounded-lg"
+		onclick={handleActivate}
+		onauxclick={(e) => onAuxClick?.(tab.id, e)}
+		aria-current={isActive ? 'page' : undefined}
+		aria-label={tab.name}
+	></a>
+
+	{#if isLoading}
+		<Tooltip.Root>
+			<Tooltip.Trigger>
+				{#snippet child({ props })}
+					<button
+						{...props}
+						class="stop-button relative z-10 flex h-5 w-5 shrink-0 cursor-pointer items-center justify-center rounded-sm text-muted-foreground transition-colors hover:text-foreground"
+						onclick={(e) => handleActionClick(e, () => onStop?.(tab.id, e))}
+						aria-label="Stop generation"
+					>
+						<Loader2
+							class="loading-icon {ICON_CLASS_SM} animate-spin transition-opacity duration-300 {contentOpacity}"
+						/>
+
+						<Square
+							class="stop-icon hidden {ICON_CLASS_XS} fill-current text-destructive transition-opacity {contentOpacity}"
+						/>
+					</button>
+				{/snippet}
+			</Tooltip.Trigger>
+
+			<Tooltip.Content>
+				<p>Stop generation</p>
+			</Tooltip.Content>
+		</Tooltip.Root>
+	{/if}
+
+	{#if tab.isNewChat}
+		<SquarePen
+			class="pointer-events-none {ICON_CLASS_SM} shrink-0 transition-opacity {contentOpacity}"
+		/>
+	{/if}
+
+	<span class="pointer-events-none truncate transition-opacity {contentOpacity}">{tab.name}</span>
+
+	<Tooltip.Root>
+		<Tooltip.Trigger>
+			{#snippet child({ props })}
+				<button
+					{...props}
+					class={cn(
+						'relative z-10 flex h-5 w-5 shrink-0 cursor-pointer items-center justify-center rounded-sm text-muted-foreground transition-opacity hover:bg-foreground/10 hover:text-foreground',
+						contentOpacity
+					)}
+					onclick={(e) => handleActionClick(e, () => onClose?.(tab.id))}
+					aria-label="Close tab"
+				>
+					<X class={ICON_CLASS_SM} />
+				</button>
+			{/snippet}
+		</Tooltip.Trigger>
+
+		<Tooltip.Content>
+			<p>Close tab</p>
+		</Tooltip.Content>
+	</Tooltip.Root>
+</div>
+
+<style>
+	.stop-button {
+		:global(.stop-icon) {
+			display: none;
+		}
+
+		:global(.loading-icon) {
+			display: block;
+		}
+
+		&:is(:hover) {
+			:global(.stop-icon) {
+				display: block;
+			}
+
+			:global(.loading-icon) {
+				display: none;
+			}
+		}
+	}
+</style>
diff --git a/tools/ui/src/lib/components/app/chat/ChatTabs/ChatTabsNewChatButton.svelte b/tools/ui/src/lib/components/app/chat/ChatTabs/ChatTabsNewChatButton.svelte
new file mode 100644
index 000000000..d62e551f7
--- /dev/null
+++ b/tools/ui/src/lib/components/app/chat/ChatTabs/ChatTabsNewChatButton.svelte
@@ -0,0 +1,30 @@
+<script lang="ts">
+	import { Plus } from '@lucide/svelte';
+	import * as Tooltip from '$lib/components/ui/tooltip';
+	import { ICON_CLASS_DEFAULT } from '$lib/constants';
+
+	interface Props {
+		onclick?: () => void;
+	}
+
+	let { onclick }: Props = $props();
+</script>
+
+<Tooltip.Root>
+	<Tooltip.Trigger>
+		{#snippet child({ props })}
+			<button
+				{...props}
+				class="backdrop-blur-lg flex h-8 w-8 mr-4 shrink-0 cursor-pointer items-center justify-center rounded-md transition-colors hover:bg-foreground/5"
+				{onclick}
+				aria-label="New chat"
+			>
+				<Plus class="{ICON_CLASS_DEFAULT} opacity-40 transition-opacity group-hover:opacity-100" />
+			</button>
+		{/snippet}
+	</Tooltip.Trigger>
+
+	<Tooltip.Content>
+		<p>New chat</p>
+	</Tooltip.Content>
+</Tooltip.Root>
diff --git a/tools/ui/src/lib/components/app/chat/index.ts b/tools/ui/src/lib/components/app/chat/index.ts
index 34571d53b..a96ae3789 100644
--- a/tools/ui/src/lib/components/app/chat/index.ts
+++ b/tools/ui/src/lib/components/app/chat/index.ts
@@ -686,6 +686,18 @@ export { default as ChatMessageSystem } from './ChatMessages/ChatMessage/ChatMes
  */
 export { default as ChatScreen } from './ChatScreen/ChatScreen.svelte';
 
+/**
+ * **ChatTabs** - Browser-style tab bar for open conversations
+ *
+ * Horizontal strip of tabs rendered above ChatScreen in the chat layout,
+ * one per conversation tracked by tabsStore. The active tab follows the
+ * route's conversation id; clicking a tab navigates to it, middle-click or
+ * the close button closes it (switching to the left neighbor when closing
+ * the active tab), and a trailing "+" button starts a new chat. Shows a
+ * spinner on tabs with a running generation. Desktop-only.
+ */
+export { default as ChatTabs } from './ChatTabs/ChatTabs.svelte';
+
 /**
  * Visual overlay displayed when user drags files over the chat screen.
  * Shows drop zone indicator to guide users where to release files.
diff --git a/tools/ui/src/lib/components/app/misc/HorizontalScrollCarousel.svelte b/tools/ui/src/lib/components/app/misc/HorizontalScrollCarousel.svelte
deleted file mode 100644
index e2edb4d02..000000000
--- a/tools/ui/src/lib/components/app/misc/HorizontalScrollCarousel.svelte
+++ /dev/null
@@ -1,96 +0,0 @@
-<script lang="ts">
-	import { ChevronLeft, ChevronRight } from '@lucide/svelte';
-	import { ICON_CLASS_DEFAULT } from '$lib/constants';
-	import type { Snippet } from 'svelte';
-
-	interface Props {
-		class?: string;
-		children?: Snippet;
-		gapSize?: string;
-		onScrollableChange?: (isScrollable: boolean) => void;
-	}
-
-	let { children, class: className = '', gapSize = '3', onScrollableChange }: Props = $props();
-
-	let canScrollLeft = $state(false);
-	let canScrollRight = $state(false);
-	let scrollContainer: HTMLDivElement | undefined = $state();
-
-	function scrollLeft(event?: MouseEvent) {
-		event?.stopPropagation();
-		event?.preventDefault();
-
-		if (!scrollContainer) return;
-
-		scrollContainer.scrollBy({ behavior: 'smooth', left: scrollContainer.clientWidth * -0.67 });
-	}
-
-	function scrollRight(event?: MouseEvent) {
-		event?.stopPropagation();
-		event?.preventDefault();
-
-		if (!scrollContainer) return;
-
-		scrollContainer.scrollBy({ behavior: 'smooth', left: scrollContainer.clientWidth * 0.67 });
-	}
-
-	function updateScrollButtons() {
-		if (!scrollContainer) return;
-
-		const { clientWidth, scrollLeft, scrollWidth } = scrollContainer;
-
-		canScrollLeft = scrollLeft > 0;
-		canScrollRight = scrollLeft < scrollWidth - clientWidth - 1;
-
-		const isScrollable = scrollWidth > clientWidth;
-
-		onScrollableChange?.(isScrollable);
-	}
-
-	export function resetScroll() {
-		if (scrollContainer) {
-			scrollContainer.scrollLeft = 0;
-			setTimeout(() => {
-				updateScrollButtons();
-			}, 0);
-		}
-	}
-
-	$effect(() => {
-		if (!scrollContainer) return;
-
-		const observer = new ResizeObserver(() => updateScrollButtons());
-
-		observer.observe(scrollContainer);
-
-		return () => observer.disconnect();
-	});
-</script>
-
-<div class="relative {className}">
-	<button
-		class="absolute top-1/2 left-4 z-10 flex h-6 w-6 -translate-y-1/2 items-center justify-center rounded-full bg-background/25 shadow-md backdrop-blur-xs transition-opacity hover:bg-background/45 disabled:pointer-events-none disabled:opacity-0"
-		onclick={scrollLeft}
-		disabled={!canScrollLeft}
-		aria-label="Scroll left"
-	>
-		<ChevronLeft class={ICON_CLASS_DEFAULT} />
-	</button>
-
-	<div
-		class="scrollbar-hide flex items-start gap-{gapSize} overflow-x-auto"
-		bind:this={scrollContainer}
-		onscroll={updateScrollButtons}
-	>
-		{@render children?.()}
-	</div>
-
-	<button
-		class="absolute top-1/2 right-4 z-10 flex h-6 w-6 -translate-y-1/2 items-center justify-center rounded-full bg-background/25 shadow-md backdrop-blur-xs transition-opacity hover:bg-background/45 disabled:pointer-events-none disabled:opacity-0"
-		onclick={scrollRight}
-		disabled={!canScrollRight}
-		aria-label="Scroll right"
-	>
-		<ChevronRight class={ICON_CLASS_DEFAULT} />
-	</button>
-</div>
diff --git a/tools/ui/src/lib/components/app/misc/ScrollCarousel.svelte b/tools/ui/src/lib/components/app/misc/ScrollCarousel.svelte
new file mode 100644
index 000000000..a0a914f5e
--- /dev/null
+++ b/tools/ui/src/lib/components/app/misc/ScrollCarousel.svelte
@@ -0,0 +1,131 @@
+<script lang="ts">
+	import { ChevronLeft, ChevronRight } from '@lucide/svelte';
+	import { cn } from '$lib/components/ui/utils';
+	import { ICON_CLASS_DEFAULT } from '$lib/constants';
+	import { ScrollCarouselVariant } from '$lib/enums';
+	import { useScrollCarousel } from '$lib/hooks/use-scroll-carousel.svelte';
+	import type { Snippet } from 'svelte';
+
+	interface Props {
+		children: Snippet;
+		/** External carousel hook for callers that need to drive it (e.g. scrollToCenter). */
+		carousel?: ReturnType<typeof useScrollCarousel>;
+		/** Classes for the outer relative wrapper. */
+		class?: string;
+		/** Classes for the scrollable overflow container. */
+		containerClass?: string;
+		/** Classes for the min-w-max content wrapper. */
+		innerClass?: string;
+		/** Tailwind gap class applied to the content wrapper. */
+		gapSize?: string;
+		/** Show the arrows whenever the content overflows, even without hover. */
+		alwaysShowArrows?: boolean;
+		/** Arrow placement and styling. */
+		variant?: ScrollCarouselVariant;
+	}
+
+	let {
+		alwaysShowArrows = false,
+		carousel: externalCarousel,
+		children,
+		class: className = '',
+		containerClass = '',
+		gapSize = '3',
+		innerClass = '',
+		variant = ScrollCarouselVariant.TOP
+	}: Props = $props();
+
+	const internalCarousel = useScrollCarousel();
+	const carousel = $derived(externalCarousel ?? internalCarousel);
+
+	const isCenter = $derived(variant === ScrollCarouselVariant.CENTER);
+
+	function scrollLeft(event?: MouseEvent) {
+		event?.stopPropagation();
+		event?.preventDefault();
+
+		const container = carousel.scrollContainer;
+
+		if (!container) return;
+
+		container.scrollBy({ behavior: 'smooth', left: -(container.clientWidth * 0.67) });
+	}
+
+	function scrollRight(event?: MouseEvent) {
+		event?.stopPropagation();
+		event?.preventDefault();
+
+		const container = carousel.scrollContainer;
+
+		if (!container) return;
+
+		container.scrollBy({ behavior: 'smooth', left: container.clientWidth * 0.67 });
+	}
+
+	export function resetScroll() {
+		const container = carousel.scrollContainer;
+
+		if (!container) return;
+
+		container.scrollLeft = 0;
+		setTimeout(() => carousel.updateScrollButtons(), 0);
+	}
+</script>
+
+<div
+	class={cn('group relative', !isCenter && 'flex items-center', className)}
+	style={!isCenter ? 'scroll-padding: 1rem;' : undefined}
+>
+	<button
+		class={cn(
+			'absolute z-10 flex h-6 w-6 items-center justify-center rounded-full shadow-md transition-opacity',
+			isCenter
+				? 'top-1/2 left-4 -translate-y-1/2 bg-background/25 backdrop-blur-xs hover:bg-background/45 disabled:pointer-events-none disabled:opacity-0'
+				: 'left-2 bg-muted backdrop-blur-sm hover:bg-accent',
+			!isCenter &&
+				(carousel.canScrollLeft
+					? alwaysShowArrows
+						? 'opacity-100'
+						: 'opacity-0 group-hover:opacity-100'
+					: 'pointer-events-none opacity-0')
+		)}
+		{...isCenter ? { disabled: !carousel.canScrollLeft } : {}}
+		onclick={scrollLeft}
+		aria-label="Scroll left"
+	>
+		<ChevronLeft class={ICON_CLASS_DEFAULT} />
+	</button>
+
+	<div
+		class={cn('scrollbar-hide overflow-x-auto', containerClass)}
+		bind:this={carousel.scrollContainer}
+		onscroll={carousel.updateScrollButtons}
+	>
+		<div
+			class={cn('flex min-w-max', isCenter && 'items-start', `gap-${gapSize}`, innerClass)}
+			bind:this={carousel.contentContainer}
+		>
+			{@render children?.()}
+		</div>
+	</div>
+
+	<button
+		class={cn(
+			'absolute z-10 flex h-6 w-6 items-center justify-center rounded-full shadow-md transition-opacity',
+			isCenter
+				? 'top-1/2 right-4 -translate-y-1/2 bg-background/25 backdrop-blur-xs hover:bg-background/45 disabled:pointer-events-none disabled:opacity-0'
+				: 'right-2 bg-muted backdrop-blur-sm hover:bg-accent',
+			!isCenter &&
+				(carousel.canScrollRight
+					? alwaysShowArrows
+						? 'opacity-100'
+						: 'opacity-0 group-hover:opacity-100'
+					: 'pointer-events-none opacity-0')
+		)}
+		{...isCenter ? { disabled: !carousel.canScrollRight } : {}}
+		onclick={scrollRight}
+		aria-label="Scroll right"
+	>
+		<ChevronRight class={ICON_CLASS_DEFAULT} />
+	</button>
+</div>
diff --git a/tools/ui/src/lib/components/app/misc/index.ts b/tools/ui/src/lib/components/app/misc/index.ts
index b550ae66a..a10410ef9 100644
--- a/tools/ui/src/lib/components/app/misc/index.ts
+++ b/tools/ui/src/lib/components/app/misc/index.ts
@@ -21,13 +21,6 @@
  */
 export { default as ConversationSelection } from './ConversationSelection.svelte';
 
-/**
- * Horizontal scrollable carousel with navigation arrows.
- * Used for displaying items in a horizontally scrollable container
- * with left/right navigation buttons that appear on hover.
- */
-export { default as HorizontalScrollCarousel } from './HorizontalScrollCarousel.svelte';
-
 /**
  * **TruncatedText** - Text with ellipsis and tooltip
  *
@@ -44,6 +37,13 @@ export { default as TruncatedText } from './TruncatedText.svelte';
  */
 export { default as KeyboardShortcutInfo } from './KeyboardShortcutInfo.svelte';
 
+/**
+ * **ScrollCarousel** - Feature/carousel with center-aligned overflow controls
+ *
+ * Horizontal scrollable container with arrows that center the focused item.
+ */
+export { default as ScrollCarousel } from './ScrollCarousel.svelte';
+
 /**
  * **CodeBlockActions** - Actions bar for code blocks (copy, preview)
  *
diff --git a/tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigation.svelte b/tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigation.svelte
index aa63c2915..424f2feca 100644
--- a/tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigation.svelte
+++ b/tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigation.svelte
@@ -14,7 +14,7 @@
 	import { useKeyboardShortcuts } from '$lib/hooks/use-keyboard-shortcuts.svelte';
 	import { useMarqueeSelection } from '$lib/hooks/use-marquee-selection.svelte';
 	import { RouterService } from '$lib/services/router.service';
-	import { chatStore, conversationsStore, deviceStore, settingsStore } from '$lib/stores';
+	import { chatStore, conversationsStore, deviceStore, settingsStore, uiStore } from '$lib/stores';
 	import { buildConversationTree } from '$lib/utils';
 	import { circIn } from 'svelte/easing';
 	import { SvelteSet } from 'svelte/reactivity';
@@ -31,30 +31,29 @@
 		toggleSidebar: () => toggleExpandedMode()
 	});
 
-	let isExpandedMode = $state(false);
 	let hoveredTooltip = $state<string | null>(null);
 	let logoHovered = $state(false);
 
-	const isStripExpanded = $derived(isExpandedMode || hoveredTooltip !== null);
+	const isStripExpanded = $derived(uiStore.isSidebarExpanded || hoveredTooltip !== null);
 	const isOnMobile = $derived(deviceStore.isMobile);
 	const alwaysShowOnDesktop = $derived(settingsStore.config.alwaysShowSidebarOnDesktop as boolean);
 
 	$effect(() => {
 		if (alwaysShowOnDesktop && !isOnMobile) {
-			isExpandedMode = true;
+			uiStore.isSidebarExpanded = true;
 		}
 	});
 
 	function toggleExpandedMode() {
-		isExpandedMode = !isExpandedMode;
+		uiStore.isSidebarExpanded = !uiStore.isSidebarExpanded;
 
-		if (!isExpandedMode) {
+		if (!uiStore.isSidebarExpanded) {
 			hoveredTooltip = null;
 		}
 	}
 
 	$effect(() => {
-		if (!isExpandedMode) {
+		if (!uiStore.isSidebarExpanded) {
 			isSearchModeActive = false;
 			searchQuery = '';
 
@@ -66,7 +65,7 @@
 
 	$effect(() => {
 		if (deviceStore.isMobile && page.url.hash.includes(ROUTES.SEARCH)) {
-			isExpandedMode = false;
+			uiStore.isSidebarExpanded = false;
 		}
 	});
 
@@ -294,7 +293,7 @@
 		}
 
 		pendingCollapse = setTimeout(() => {
-			isExpandedMode = false;
+			uiStore.isSidebarExpanded = false;
 			pendingCollapse = null;
 		}, 100);
 	}
@@ -314,7 +313,7 @@
 		class={[
 			'fixed md:sticky top-2 left-2 md:left-0 md:ml-2 md:mt-2 pt-2 z-10 w-[calc(100dvw-1rem)]',
 			'md:h-[calc(100dvh-1.125rem)]',
-			isExpandedMode &&
+			uiStore.isSidebarExpanded &&
 				(deviceStore.isStandalone
 					? 'h-[calc(100dvh-2rem)]'
 					: deviceStore.isIOSDevice
@@ -323,9 +322,9 @@
 			'rounded-3xl md:rounded-2xl',
 			'flex flex-col justify-between',
 			'md:transition-[width,padding] duration-200 ease-out',
-			isStripExpanded && 'md:w-72 md:bg-muted/60 md:backdrop-blur-xl border-border shadow-md',
+			isStripExpanded && 'md:w-72 md:bg-muted/60 md:backdrop-blur-xl shadow-md',
 			!isStripExpanded && 'md:w-12',
-			isExpandedMode && 'is-expanded'
+			uiStore.isSidebarExpanded && 'is-expanded'
 		]}
 	>
 		<div class="px-2 flex items-center justify-between">
@@ -337,24 +336,26 @@
 				onmouseleave={() => (logoHovered = false)}
 			>
 				<ActionIcon
-					icon={!isExpandedMode && logoHovered && innerWidth > 768 ? PanelLeftOpen : Logo}
+					icon={!uiStore.isSidebarExpanded && logoHovered && innerWidth > 768
+						? PanelLeftOpen
+						: Logo}
 					size="lg"
 					iconSize="h-4.5 w-4.5 md:h-4 md:w-4"
-					class="{isExpandedMode
+					class="{uiStore.isSidebarExpanded
 						? 'bg-muted! md:bg-foreground/5!'
 						: 'bg-transparent!'} md:h-9 md:w-9 h-10 w-10 rounded-full md:hover:bg-foreground/10! pointer-events-auto"
-					href={isExpandedMode ? ROUTES.START : undefined}
-					onclick={isExpandedMode ? undefined : toggleExpandedMode}
-					tooltip={isExpandedMode ? undefined : 'Open Sidebar'}
+					href={uiStore.isSidebarExpanded ? ROUTES.START : undefined}
+					onclick={uiStore.isSidebarExpanded ? undefined : toggleExpandedMode}
+					tooltip={uiStore.isSidebarExpanded ? undefined : 'Open Sidebar'}
 					tooltipSide={TooltipSide.RIGHT}
-					ariaLabel={isExpandedMode ? 'Go to start' : 'Expand navigation'}
+					ariaLabel={uiStore.isSidebarExpanded ? 'Go to start' : 'Expand navigation'}
 				/>
 			</div>
 
-			{#if isOnMobile || (isExpandedMode && !alwaysShowOnDesktop)}
+			{#if isOnMobile || (uiStore.isSidebarExpanded && !alwaysShowOnDesktop)}
 				<div
 					class="flex items-center transition-all duration-150 ease-out {deviceStore.isMobile &&
-					!isExpandedMode
+					!uiStore.isSidebarExpanded
 						? 'opacity-0 h-0!'
 						: ''}"
 					in:fade={{ delay: 50, duration: 150, easing: circIn }}
@@ -377,12 +378,12 @@
 		<div
 			class="mt-2 flex min-h-0 flex-1 flex-col gap-4 md:gap-1 {deviceStore.isMobile
 				? 'transition-[opacity,height] duration-200 ease-out'
-				: ''} {deviceStore.isMobile && !isExpandedMode ? 'opacity-0 !h-0' : ''}"
+				: ''} {deviceStore.isMobile && !uiStore.isSidebarExpanded ? 'opacity-0 !h-0' : ''}"
 			in:fade={{ duration: 200 }}
 			out:fade={{ duration: 200 }}
 		>
 			<SidebarNavigationActions
-				isExpandedMode={innerWidth > 768 ? isExpandedMode : true}
+				isExpandedMode={innerWidth > 768 ? uiStore.isSidebarExpanded : true}
 				class="px-2"
 				bind:isSearchModeActive
 				bind:searchQuery
@@ -391,7 +392,7 @@
 					searchQuery = '';
 				}}
 				onSearchClick={() => {
-					isExpandedMode = true;
+					uiStore.isSidebarExpanded = true;
 					isSearchModeActive = true;
 				}}
 				onNewChat={() => {
@@ -401,7 +402,7 @@
 				}}
 			/>
 
-			{#if isExpandedMode || isOnMobile}
+			{#if uiStore.isSidebarExpanded || isOnMobile}
 				<div class="flex min-h-0 flex-1 flex-col overflow-y-auto">
 					<SidebarNavigationConversationList
 						class="px-2"
diff --git a/tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigationActions.svelte b/tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigationActions.svelte
index dc333134b..42a0be707 100644
--- a/tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigationActions.svelte
+++ b/tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigationActions.svelte
@@ -11,8 +11,8 @@
 		ROUTES,
 		SIDEBAR_ACTIONS_ITEMS
 	} from '$lib/constants';
-	import { TooltipSide } from '$lib/enums';
-	import { deviceStore } from '$lib/stores';
+	import { SidebarAction, TooltipSide } from '$lib/enums';
+	import { conversationsStore, deviceStore } from '$lib/stores';
 	import type { Component } from 'svelte';
 	import { onMount } from 'svelte';
 	import { circIn } from 'svelte/easing';
@@ -109,14 +109,20 @@
 			{@const isActive = isItemActive(item)}
 			{@const isSearchOnMobile = item.icon === Search && deviceStore.isMobile}
 			{@const itemHref = isSearchOnMobile ? ROUTES.SEARCH : item.route}
-			{@const itemOnClick = item.route
-				? () => {
-						onNewChat?.();
-						goto(item.route!);
-					}
-				: isSearchOnMobile
-					? undefined
-					: onSearchClick}
+			{@const itemOnClick =
+				item.action === SidebarAction.NEW_CHAT
+					? () => {
+							onNewChat?.();
+							void conversationsStore.openNewChat();
+						}
+					: item.route
+						? () => {
+								onNewChat?.();
+								goto(item.route!);
+							}
+						: isSearchOnMobile
+							? undefined
+							: onSearchClick}
 			{@const itemTransition = {
 				delay: !initialized ? i * ICON_STRIP_TRANSITION_DELAY_MULTIPLIER : 0,
 				duration: ICON_STRIP_TRANSITION_DURATION,
@@ -157,14 +163,20 @@
 		{#each SIDEBAR_ACTIONS_ITEMS as item, i (item.tooltip)}
 			{@const isActive = isItemActive(item)}
 			{@const isSearchOnMobile = item.icon === Search && deviceStore.isMobile}
-			{@const itemOnClick = item.route
-				? () => {
-						onNewChat?.();
-						goto(item.route!);
-					}
-				: isSearchOnMobile
-					? undefined
-					: onSearchClick}
+			{@const itemOnClick =
+				item.action === SidebarAction.NEW_CHAT
+					? () => {
+							onNewChat?.();
+							void conversationsStore.openNewChat();
+						}
+					: item.route
+						? () => {
+								onNewChat?.();
+								goto(item.route!);
+							}
+						: isSearchOnMobile
+							? undefined
+							: onSearchClick}
 			{@const itemTransition = {
 				delay: !initialized ? i * ICON_STRIP_TRANSITION_DELAY_MULTIPLIER : 0,
 				duration: ICON_STRIP_TRANSITION_DURATION,
diff --git a/tools/ui/src/lib/components/app/settings/SettingsChatMobileHeader.svelte b/tools/ui/src/lib/components/app/settings/SettingsChatMobileHeader.svelte
index 58617956a..1e8c76f8e 100644
--- a/tools/ui/src/lib/components/app/settings/SettingsChatMobileHeader.svelte
+++ b/tools/ui/src/lib/components/app/settings/SettingsChatMobileHeader.svelte
@@ -1,5 +1,6 @@
 <script lang="ts">
-	import { ChevronLeft, ChevronRight, Settings } from '@lucide/svelte';
+	import { Settings } from '@lucide/svelte';
+	import { ScrollCarousel } from '$lib/components/app';
 	import { ICON_CLASS_DEFAULT, UI_DATA_ATTRS } from '$lib/constants';
 	import { BooleanString } from '$lib/enums';
 	import { useScrollCarousel } from '$lib/hooks/use-scroll-carousel.svelte';
@@ -44,70 +45,42 @@
 	</div>
 
 	<div class="border-b border-border/30 py-2">
-		<div class="relative flex items-center" style="scroll-padding: 1rem;">
-			<button
-				class="absolute left-2 z-10 flex h-6 w-6 items-center justify-center rounded-full bg-muted shadow-md backdrop-blur-sm transition-opacity hover:bg-accent {carousel.canScrollLeft
-					? 'opacity-100'
-					: 'pointer-events-none opacity-0'}"
-				onclick={carousel.scrollLeft}
-				aria-label="Scroll left"
-			>
-				<ChevronLeft class={ICON_CLASS_DEFAULT} />
-			</button>
-
-			<div
-				class="scrollbar-hide overflow-x-auto py-2"
-				bind:this={carousel.scrollContainer}
-				onscroll={carousel.updateScrollButtons}
-			>
-				<div class="flex min-w-max gap-2">
-					{#each sections as section (section.title)}
-						{#if getHref}
-							<a
-								class="flex cursor-pointer items-center gap-2 rounded-lg px-3 py-2 text-sm whitespace-nowrap no-underline transition-colors first:ml-4 last:mr-4 hover:bg-accent {isActive(
-									section
-								)
-									? 'bg-accent text-accent-foreground'
-									: 'text-muted-foreground'}"
-								{...{ [UI_DATA_ATTRS.ACTIVE]: isActive(section) }}
-								href={getHref(section)}
-								onclick={(e: MouseEvent) => {
-									carousel.scrollToCenter(e.currentTarget as HTMLElement);
-								}}
-							>
-								<section.icon class="{ICON_CLASS_DEFAULT} flex-shrink-0" />
-								<span>{section.title}</span>
-							</a>
-						{:else}
-							<button
-								class="flex cursor-pointer items-center gap-2 rounded-lg px-3 py-2 text-sm whitespace-nowrap transition-colors first:ml-4 last:mr-4 hover:bg-accent {isActive(
-									section
-								)
-									? 'bg-accent text-accent-foreground'
-									: 'text-muted-foreground'}"
-								{...{ [UI_DATA_ATTRS.ACTIVE]: isActive(section) }}
-								onclick={(e: MouseEvent) => {
-									onSectionChange?.(section.title);
-									carousel.scrollToCenter(e.currentTarget as HTMLElement);
-								}}
-							>
-								<section.icon class="{ICON_CLASS_DEFAULT} flex-shrink-0" />
-								<span>{section.title}</span>
-							</button>
-						{/if}
-					{/each}
-				</div>
-			</div>
-
-			<button
-				class="absolute right-2 z-10 flex h-6 w-6 items-center justify-center rounded-full bg-muted shadow-md backdrop-blur-sm transition-opacity hover:bg-accent {carousel.canScrollRight
-					? 'opacity-100'
-					: 'pointer-events-none opacity-0'}"
-				onclick={carousel.scrollRight}
-				aria-label="Scroll right"
-			>
-				<ChevronRight class={ICON_CLASS_DEFAULT} />
-			</button>
-		</div>
+		<ScrollCarousel {carousel} alwaysShowArrows containerClass="py-2" innerClass="gap-2">
+			{#each sections as section (section.title)}
+				{#if getHref}
+					<a
+						class="flex cursor-pointer items-center gap-2 rounded-lg px-3 py-2 text-sm whitespace-nowrap no-underline transition-colors first:ml-4 last:mr-4 hover:bg-accent {isActive(
+							section
+						)
+							? 'bg-accent text-accent-foreground'
+							: 'text-muted-foreground'}"
+						{...{ [UI_DATA_ATTRS.ACTIVE]: isActive(section) }}
+						href={getHref(section)}
+						onclick={(e: MouseEvent) => {
+							carousel.scrollToCenter(e.currentTarget as HTMLElement);
+						}}
+					>
+						<section.icon class="{ICON_CLASS_DEFAULT} flex-shrink-0" />
+						<span>{section.title}</span>
+					</a>
+				{:else}
+					<button
+						class="flex cursor-pointer items-center gap-2 rounded-lg px-3 py-2 text-sm whitespace-nowrap transition-colors first:ml-4 last:mr-4 hover:bg-accent {isActive(
+							section
+						)
+							? 'bg-accent text-accent-foreground'
+							: 'text-muted-foreground'}"
+						{...{ [UI_DATA_ATTRS.ACTIVE]: isActive(section) }}
+						onclick={(e: MouseEvent) => {
+							onSectionChange?.(section.title);
+							carousel.scrollToCenter(e.currentTarget as HTMLElement);
+						}}
+					>
+						<section.icon class="{ICON_CLASS_DEFAULT} flex-shrink-0" />
+						<span>{section.title}</span>
+					</button>
+				{/if}
+			{/each}
+		</ScrollCarousel>
 	</div>
 </div>
diff --git a/tools/ui/src/lib/constants/chat-tabs.constants.ts b/tools/ui/src/lib/constants/chat-tabs.constants.ts
new file mode 100644
index 000000000..1c0bb051a
--- /dev/null
+++ b/tools/ui/src/lib/constants/chat-tabs.constants.ts
@@ -0,0 +1,18 @@
+/** Sentinel tab id for the bare `#/` new-chat screen */
+export const NEW_CHAT_TAB_ID = 'new-chat';
+
+/** Label shown for the new-chat sentinel tab. */
+export const NEW_CHAT_LABEL = 'New chat';
+
+/** Fallback label for conversations without an auto-generated title. */
+export const UNNAMED_CHAT_LABEL = 'Chat';
+
+/**
+ * Tab bar max width so it stays clear of the sidebar strip. The expanded strip
+ * is `md:w-72` and the collapsed one `md:w-12`; these hold the fully tuned
+ * `max-w-[calc(100vw-?rem)]` classes so the offset has a single source.
+ */
+export const CHAT_TABS_MAX_WIDTH = {
+	COLLAPSED_SIDEBAR: 'max-w-[calc(100vw-5rem)]',
+	EXPANDED_SIDEBAR: 'max-w-[calc(100vw-20rem)]'
+} as const;
diff --git a/tools/ui/src/lib/constants/css-classes.constants.ts b/tools/ui/src/lib/constants/css-classes.constants.ts
index 4e3310544..07dd77ff5 100644
--- a/tools/ui/src/lib/constants/css-classes.constants.ts
+++ b/tools/ui/src/lib/constants/css-classes.constants.ts
@@ -26,5 +26,11 @@ export const CHAT_INPUT_FOCUS_SELECTOR =
 /** Default Tailwind size class for inline icon components (lucide, etc.). */
 export const ICON_CLASS_DEFAULT = 'h-4 w-4';
 
+/** Small Tailwind size class for inline icons. */
+export const ICON_CLASS_SM = 'h-3.5 w-3.5';
+
+/** Extra-small Tailwind size class for inline icons. */
+export const ICON_CLASS_XS = 'h-3 w-3';
+
 /** Icon size + spinning animation; used for live-streaming tool indicators. */
 export const ICON_CLASS_SPIN = 'h-4 w-4 animate-spin';
diff --git a/tools/ui/src/lib/constants/index.ts b/tools/ui/src/lib/constants/index.ts
index 3b60b2168..e3241373e 100644
--- a/tools/ui/src/lib/constants/index.ts
+++ b/tools/ui/src/lib/constants/index.ts
@@ -4,6 +4,7 @@
 export * from './agentic.constants';
 export * from './api-endpoints.constants';
 export * from './app.constants';
+export * from './chat-tabs.constants';
 export * from './database.constants';
 export * from './reasoning-effort.constants';
 export * from './recommended-mcp-servers.constants';
diff --git a/tools/ui/src/lib/constants/routes.constants.ts b/tools/ui/src/lib/constants/routes.constants.ts
index e4cce7739..d46acfd78 100644
--- a/tools/ui/src/lib/constants/routes.constants.ts
+++ b/tools/ui/src/lib/constants/routes.constants.ts
@@ -4,8 +4,6 @@ export const URL_PARAMS = {
 	LOAD: 'load',
 	/** Model to select. */
 	MODEL: 'model',
-	/** Start a new chat. */
-	NEW_CHAT: 'new_chat',
 	/** Prompt to send on arrival. */
 	QUERY: 'q'
 } as const;
@@ -15,8 +13,6 @@ export const ROUTES = {
 	CHAT: '#/chat',
 	/** MCP servers. */
 	MCP_SERVERS: '#/mcp-servers',
-	/** New chat — root with new chat query param. */
-	NEW_CHAT: `?${URL_PARAMS.NEW_CHAT}=true#/`,
 	/** Search — mobile-only full-page conversation search. */
 	SEARCH: '#/search',
 	/** Settings base — for dynamic settings URLs use RouterService. */
diff --git a/tools/ui/src/lib/constants/settings-keys.constants.ts b/tools/ui/src/lib/constants/settings-keys.constants.ts
index b53d11048..c8761c758 100644
--- a/tools/ui/src/lib/constants/settings-keys.constants.ts
+++ b/tools/ui/src/lib/constants/settings-keys.constants.ts
@@ -11,6 +11,7 @@ export const SETTINGS_KEYS = {
 	API_KEY: 'apiKey',
 	AUTO_MIC_ON_EMPTY: 'autoMicOnEmpty',
 	BACKEND_SAMPLING: 'backend_sampling',
+	CONVERSATION_TABS: 'conversationTabs',
 	COPY_TEXT_ATTACHMENTS_AS_PLAIN_TEXT: 'copyTextAttachmentsAsPlainText',
 	CUSTOM_CSS: 'customCss',
 	// PY_INTERPRETER_ENABLED: 'pyInterpreterEnabled',
diff --git a/tools/ui/src/lib/constants/settings.constants.ts b/tools/ui/src/lib/constants/settings.constants.ts
index e5789c528..4d7807547 100644
--- a/tools/ui/src/lib/constants/settings.constants.ts
+++ b/tools/ui/src/lib/constants/settings.constants.ts
@@ -255,6 +255,13 @@ export const SETTINGS_REGISTRY: SettingsSectionEntry[] = [
 				label: 'Always show sidebar on desktop',
 				type: SettingsFieldType.CHECKBOX
 			},
+			{
+				defaultValue: true,
+				help: 'Show open chats as browser-style tabs above the conversation, one per open chat. When disabled, only one chat is shown at a time.',
+				key: SETTINGS_KEYS.CONVERSATION_TABS,
+				label: 'Conversation tabs',
+				type: SettingsFieldType.CHECKBOX
+			},
 			{
 				defaultValue: false,
 				help: 'Display full raw model identifiers (e.g. "ggml-org/GLM-4.7-Flash-GGUF:Q8_0") instead of parsed names with badges.',
diff --git a/tools/ui/src/lib/constants/storage.constants.ts b/tools/ui/src/lib/constants/storage.constants.ts
index 5d9acaafb..918ee4508 100644
--- a/tools/ui/src/lib/constants/storage.constants.ts
+++ b/tools/ui/src/lib/constants/storage.constants.ts
@@ -22,6 +22,7 @@ export const DISABLED_TOOLS_LOCALSTORAGE_KEY = `${STORAGE_APP_NAME}.disabledTool
 export const DISABLED_TOOL_KEYS_LOCALSTORAGE_KEY = `${STORAGE_APP_NAME}.disabledToolKeys`;
 export const FAVORITE_MODELS_LOCALSTORAGE_KEY = `${STORAGE_APP_NAME}.favoriteModels`;
 export const REASONING_EFFORT_DEFAULT_LOCALSTORAGE_KEY = `${STORAGE_APP_NAME}.reasoningEffortDefault`;
+export const CONVERSATION_TABS_LOCALSTORAGE_KEY = `${STORAGE_APP_NAME}.conversationTabs`;
 export const USER_OVERRIDES_LOCALSTORAGE_KEY = `${STORAGE_APP_NAME}.userOverrides`;
 export const DISMISSED_RECOMMENDED_MCP_SERVERS_LOCALSTORAGE_KEY = `${STORAGE_APP_NAME}.dismissedRecommendedMcpServers`;
 
diff --git a/tools/ui/src/lib/constants/ui.constants.ts b/tools/ui/src/lib/constants/ui.constants.ts
index feae08742..0f56b8d56 100644
--- a/tools/ui/src/lib/constants/ui.constants.ts
+++ b/tools/ui/src/lib/constants/ui.constants.ts
@@ -1,7 +1,7 @@
 import { ROUTES } from './routes.constants';
 import { Package, Search, Settings, SquarePen } from '@lucide/svelte';
 import McpLogo from '$lib/components/app/mcp/McpLogo.svelte';
-import { ToolSource } from '$lib/enums/tools.enums';
+import { SidebarAction, ToolSource } from '$lib/enums';
 import type { DesktopIconStripItem } from '$lib/types';
 
 export const FORK_TREE_DEPTH_PADDING = 8;
@@ -10,6 +10,7 @@ export const SYSTEM_MESSAGE_PLACEHOLDER = 'System message';
 /** Data attributes for app-level DOM contracts. */
 export const UI_DATA_ATTRS = {
 	ACTIVE: 'data-active',
+	ACTIVE_TAB: 'data-active-tab',
 	CONVERSATION_ROW: 'data-conversation-row',
 	HIGHLIGHT_THEME_PREVIEW: 'data-highlight-theme-preview',
 	PICKER_INDEX: 'data-picker-index',
@@ -55,7 +56,12 @@ export const ICON_STRIP_TRANSITION_DELAY_MULTIPLIER = 50;
 export const MAX_HEIGHT_CODE_BLOCK = '22rem';
 
 export const SIDEBAR_ACTIONS_ITEMS: DesktopIconStripItem[] = [
-	{ icon: SquarePen, keys: ['shift', 'cmd', 'o'], route: ROUTES.NEW_CHAT, tooltip: 'New chat' },
+	{
+		action: SidebarAction.NEW_CHAT,
+		icon: SquarePen,
+		keys: ['shift', 'cmd', 'o'],
+		tooltip: 'New chat'
+	},
 	{ icon: Search, keys: ['cmd', 'k'], tooltip: 'Search' },
 	{
 		activeRouteId: '/mcp-servers',
diff --git a/tools/ui/src/lib/enums/index.ts b/tools/ui/src/lib/enums/index.ts
index 7ad7df7b8..89105c6b0 100644
--- a/tools/ui/src/lib/enums/index.ts
+++ b/tools/ui/src/lib/enums/index.ts
@@ -79,6 +79,8 @@ export {
 	ColorMode,
 	HtmlInputType,
 	McpPromptVariant,
+	ScrollCarouselVariant,
+	SidebarAction,
 	TooltipSide,
 	UrlProtocol
 } from './ui.enums';
diff --git a/tools/ui/src/lib/enums/keyboard.enums.ts b/tools/ui/src/lib/enums/keyboard.enums.ts
index fb47fcf3e..3fde816f6 100644
--- a/tools/ui/src/lib/enums/keyboard.enums.ts
+++ b/tools/ui/src/lib/enums/keyboard.enums.ts
@@ -7,6 +7,8 @@ export enum KeyboardKey {
 	ARROW_RIGHT = 'ArrowRight',
 	ARROW_UP = 'ArrowUp',
 	B_LOWER = 'b',
+	BRACKET_LEFT = 'BracketLeft',
+	BRACKET_RIGHT = 'BracketRight',
 	D_LOWER = 'd',
 	D_UPPER = 'D',
 	E_UPPER = 'E',
@@ -16,5 +18,7 @@ export enum KeyboardKey {
 	O_LOWER = 'o',
 	O_UPPER = 'O',
 	SPACE = ' ',
-	TAB = 'Tab'
+	TAB = 'Tab',
+	X_LOWER = 'x',
+	X_UPPER = 'X'
 }
diff --git a/tools/ui/src/lib/enums/ui.enums.ts b/tools/ui/src/lib/enums/ui.enums.ts
index 0de34ccfa..34bbf72c5 100644
--- a/tools/ui/src/lib/enums/ui.enums.ts
+++ b/tools/ui/src/lib/enums/ui.enums.ts
@@ -11,6 +11,21 @@ export enum TooltipSide {
 	TOP = 'top'
 }
 
+/**
+ * ScrollCarousel arrow placement.
+ */
+export enum ScrollCarouselVariant {
+	CENTER = 'center',
+	TOP = 'top'
+}
+
+/**
+ * Sidebar icon strip actions handled directly by the sidebar.
+ */
+export enum SidebarAction {
+	NEW_CHAT = 'new-chat'
+}
+
 /**
  * MCP prompt display variant
  */
diff --git a/tools/ui/src/lib/hooks/use-keyboard-shortcuts.svelte.ts b/tools/ui/src/lib/hooks/use-keyboard-shortcuts.svelte.ts
index ce394c383..eef1bc332 100644
--- a/tools/ui/src/lib/hooks/use-keyboard-shortcuts.svelte.ts
+++ b/tools/ui/src/lib/hooks/use-keyboard-shortcuts.svelte.ts
@@ -1,6 +1,7 @@
-import { goto } from '$app/navigation';
-import { ROUTES } from '$lib/constants';
+import { page } from '$app/state';
+import { NEW_CHAT_TAB_ID } from '$lib/constants';
 import { KeyboardKey } from '$lib/enums';
+import { conversationsStore, settingsStore, tabsStore } from '$lib/stores';
 
 interface KeyboardShortcutsCallbacks {
 	activateSearchMode?: () => void;
@@ -9,6 +10,8 @@ interface KeyboardShortcutsCallbacks {
 	deleteActiveConversation?: () => void;
 	navigateToPrevConversation?: () => void;
 	navigateToNextConversation?: () => void;
+	navigateToPrevTab?: () => void;
+	navigateToNextTab?: () => void;
 	toggleSidebar?: () => void;
 }
 
@@ -34,7 +37,7 @@ export function useKeyboardShortcuts(callbacks: KeyboardShortcutsCallbacks) {
 		) {
 			event.preventDefault();
 
-			goto(ROUTES.NEW_CHAT);
+			void conversationsStore.openNewChat();
 		}
 
 		if (event.shiftKey && isCmdOrCtrl && event.key === KeyboardKey.E_UPPER) {
@@ -42,6 +45,28 @@ export function useKeyboardShortcuts(callbacks: KeyboardShortcutsCallbacks) {
 			callbacks.editActiveConversation?.();
 		}
 
+		if (
+			event.shiftKey &&
+			isCmdOrCtrl &&
+			(event.key === KeyboardKey.X_LOWER || event.key === KeyboardKey.X_UPPER)
+		) {
+			// several components register this shortcut; only let the first handler
+			// act so the synchronous navigation does not cascade-close every tab
+			if (event.defaultPrevented) return;
+
+			// close-tab only makes sense with conversation tabs enabled
+			if (!settingsStore.config.conversationTabs) return;
+
+			event.preventDefault();
+
+			const activeId =
+				page.params.id ?? (page.route.id === '/(chat)' ? NEW_CHAT_TAB_ID : undefined);
+
+			if (activeId) {
+				void tabsStore.close(activeId, activeId);
+			}
+		}
+
 		if (
 			isCmdOrCtrl &&
 			event.shiftKey &&
@@ -60,6 +85,16 @@ export function useKeyboardShortcuts(callbacks: KeyboardShortcutsCallbacks) {
 			event.preventDefault();
 			callbacks.navigateToNextConversation?.();
 		}
+
+		if (isCmdOrCtrl && event.altKey && event.shiftKey && event.code === KeyboardKey.BRACKET_LEFT) {
+			event.preventDefault();
+			callbacks.navigateToPrevTab?.();
+		}
+
+		if (isCmdOrCtrl && event.altKey && event.shiftKey && event.code === KeyboardKey.BRACKET_RIGHT) {
+			event.preventDefault();
+			callbacks.navigateToNextTab?.();
+		}
 	}
 
 	return { handleKeydown };
diff --git a/tools/ui/src/lib/hooks/use-scroll-carousel.svelte.ts b/tools/ui/src/lib/hooks/use-scroll-carousel.svelte.ts
index d23b8536b..252ceb9e3 100644
--- a/tools/ui/src/lib/hooks/use-scroll-carousel.svelte.ts
+++ b/tools/ui/src/lib/hooks/use-scroll-carousel.svelte.ts
@@ -2,6 +2,7 @@ export function useScrollCarousel() {
 	let canScrollLeft = $state(false);
 	let canScrollRight = $state(false);
 	let scrollContainer = $state<HTMLDivElement | undefined>();
+	let contentContainer = $state<HTMLDivElement | undefined>();
 
 	function scrollToCenter(element: HTMLElement) {
 		if (!scrollContainer) return;
@@ -15,18 +16,6 @@ export function useScrollCarousel() {
 		scrollContainer.scrollBy({ behavior: 'smooth', left: scrollOffset });
 	}
 
-	function scrollLeft() {
-		if (!scrollContainer) return;
-
-		scrollContainer.scrollBy({ behavior: 'smooth', left: -250 });
-	}
-
-	function scrollRight() {
-		if (!scrollContainer) return;
-
-		scrollContainer.scrollBy({ behavior: 'smooth', left: 250 });
-	}
-
 	function updateScrollButtons() {
 		if (!scrollContainer) return;
 
@@ -36,10 +25,21 @@ export function useScrollCarousel() {
 		canScrollRight = sl < scrollWidth - clientWidth - 1;
 	}
 
+	// Re-evaluate arrow visibility whenever the container or its content resizes,
+	// otherwise the arrows may not appear when overflowing items are added (e.g. new
+	// tabs/attachments) and the user has not scrolled yet.
 	$effect(() => {
-		if (scrollContainer) {
-			updateScrollButtons();
-		}
+		if (!scrollContainer) return;
+
+		updateScrollButtons();
+
+		const observer = new ResizeObserver(() => updateScrollButtons());
+
+		observer.observe(scrollContainer);
+
+		if (contentContainer) observer.observe(contentContainer);
+
+		return () => observer.disconnect();
 	});
 
 	return {
@@ -49,14 +49,18 @@ export function useScrollCarousel() {
 		get canScrollRight() {
 			return canScrollRight;
 		},
+		get contentContainer() {
+			return contentContainer;
+		},
+		set contentContainer(el: HTMLDivElement | undefined) {
+			contentContainer = el;
+		},
 		get scrollContainer() {
 			return scrollContainer;
 		},
 		set scrollContainer(el: HTMLDivElement | undefined) {
 			scrollContainer = el;
 		},
-		scrollLeft,
-		scrollRight,
 		scrollToCenter,
 		updateScrollButtons
 	};
diff --git a/tools/ui/src/lib/services/index.ts b/tools/ui/src/lib/services/index.ts
index 88e79446e..317500566 100644
--- a/tools/ui/src/lib/services/index.ts
+++ b/tools/ui/src/lib/services/index.ts
@@ -294,7 +294,7 @@ export { SandboxService } from './sandbox.service';
  * **RouterService** — Dynamic route URL construction utility
  *
  * Stateless utility for building dynamic route URLs from ROUTES base paths.
- * Static routes (START, NEW_CHAT, MCP_SERVERS) live in ROUTES constants;
+ * Static routes (START, MCP_SERVERS) live in ROUTES constants;
  * dynamic routes (CHAT, SETTINGS) are constructed here by appending parameters.
  *
  * **Architecture & Relationships:**
diff --git a/tools/ui/src/lib/stores/conversations/index.svelte.ts b/tools/ui/src/lib/stores/conversations/index.svelte.ts
index 7d6dc326c..98bf6a031 100644
--- a/tools/ui/src/lib/stores/conversations/index.svelte.ts
+++ b/tools/ui/src/lib/stores/conversations/index.svelte.ts
@@ -21,6 +21,7 @@ import {
 	type ConversationsPreferencesHost
 } from '$lib/stores/conversations/preferences.svelte';
 import { settingsStore } from '$lib/stores/settings/index.svelte';
+import { tabsStore } from '$lib/stores/tabs.svelte';
 import { filterByLeafNodeId, findLeafNode, generateConversationTitle } from '$lib/utils';
 import { SvelteSet } from 'svelte/reactivity';
 import { toast } from 'svelte-sonner';
@@ -105,8 +106,8 @@ class ConversationsStore implements ConversationsPreferencesHost {
 
 	/**
 	 * Deletes multiple conversations in sequence.
-	 * Mirrors deleteConversation() per-id; navigates to NEW_CHAT only if the
-	 * currently-open chat was among the deleted ones.
+	 * Mirrors deleteConversation() per-id; navigates to the new-chat screen only
+	 * if the currently-open chat was among the deleted ones.
 	 * @param convIds - Conversation IDs to delete
 	 */
 	async bulkDeleteConversations(convIds: string[]): Promise<void> {
@@ -138,8 +139,13 @@ class ConversationsStore implements ConversationsPreferencesHost {
 			this.notifyConversationsDeleted([...idsToRemove]);
 
 			if (activeWasDeleted) {
+				const activeId = this.activeConversation!.id;
+
+				tabsStore.removeTabs([...idsToRemove].filter((id) => id !== activeId));
 				this.clearActiveConversation();
-				await goto(ROUTES.NEW_CHAT);
+				await tabsStore.close(activeId, activeId);
+			} else {
+				tabsStore.removeTabs([...idsToRemove]);
 			}
 
 			toast.success(
@@ -276,11 +282,12 @@ class ConversationsStore implements ConversationsPreferencesHost {
 
 			this.clearActiveConversation();
 			this.conversations = [];
+			tabsStore.clear();
 			this.notifyConversationsDeleted(allIds);
 
 			toast.success('All conversations deleted');
 
-			await goto(ROUTES.NEW_CHAT);
+			await goto(ROUTES.START);
 		} catch (error) {
 			console.error('Failed to delete all conversations:', error);
 			toast.error('Failed to delete conversations');
@@ -313,8 +320,13 @@ class ConversationsStore implements ConversationsPreferencesHost {
 				this.conversations = this.conversations.filter((c) => !idsToRemove.has(c.id));
 
 				if (this.activeConversation && idsToRemove.has(this.activeConversation.id)) {
+					const activeId = this.activeConversation.id;
+
+					tabsStore.removeTabs([...idsToRemove].filter((id) => id !== activeId));
 					this.clearActiveConversation();
-					await goto(ROUTES.NEW_CHAT);
+					await tabsStore.close(activeId, activeId);
+				} else {
+					tabsStore.removeTabs([...idsToRemove]);
 				}
 
 				this.notifyConversationsDeleted([...idsToRemove]);
@@ -333,7 +345,9 @@ class ConversationsStore implements ConversationsPreferencesHost {
 
 				if (this.activeConversation?.id === convId) {
 					this.clearActiveConversation();
-					await goto(ROUTES.NEW_CHAT);
+					await tabsStore.close(convId, convId);
+				} else {
+					tabsStore.removeTabs([convId]);
 				}
 
 				this.notifyConversationsDeleted([convId]);
@@ -571,6 +585,15 @@ class ConversationsStore implements ConversationsPreferencesHost {
 		return () => this.conversationDeletionListeners.delete(listener);
 	}
 
+	/**
+	 * Start a fresh chat by navigating to the bare `#/` new-chat screen. The
+	 * chat layout opens a new-chat tab for it when Conversation tabs are on.
+	 */
+	async openNewChat(): Promise<void> {
+		this.clearActiveConversation();
+		await goto(ROUTES.START);
+	}
+
 	/**
 	 * Refreshes active messages based on currNode after branch navigation.
 	 */
@@ -676,14 +699,6 @@ class ConversationsStore implements ConversationsPreferencesHost {
 		);
 	}
 
-	/**
-	 *
-	 *
-	 * Import & Export
-	 *
-	 *
-	 */
-
 	/**
 	 * Updates the current node of the active conversation
 	 * @param nodeId - The new current node ID
@@ -717,6 +732,14 @@ class ConversationsStore implements ConversationsPreferencesHost {
 		}
 	}
 
+	/**
+	 *
+	 *
+	 * Import & Export
+	 *
+	 *
+	 */
+
 	private notifyConversationsDeleted(convIds: string[]): void {
 		if (convIds.length === 0) return;
 
diff --git a/tools/ui/src/lib/stores/conversations/preferences.svelte.ts b/tools/ui/src/lib/stores/conversations/preferences.svelte.ts
index 65a02344b..fea928603 100644
--- a/tools/ui/src/lib/stores/conversations/preferences.svelte.ts
+++ b/tools/ui/src/lib/stores/conversations/preferences.svelte.ts
@@ -152,11 +152,13 @@ export class ConversationPreferences {
 			return;
 		}
 
-		this.host.applyConversationUpdate(this.host.activeConversation.id, {
+		const id = this.host.activeConversation.id;
+
+		this.host.applyConversationUpdate(id, {
 			cwd: trimmed
 		});
 
-		await DatabaseService.updateConversation(this.host.activeConversation.id, {
+		await DatabaseService.updateConversation(id, {
 			cwd: trimmed
 		});
 
@@ -202,12 +204,15 @@ export class ConversationPreferences {
 			}
 		}
 
-		await DatabaseService.updateConversation(this.host.activeConversation.id, {
-			mcpServerOverrides: newOverrides.length > 0 ? newOverrides : undefined
+		const overrides = newOverrides.length > 0 ? newOverrides : undefined;
+		const id = this.host.activeConversation.id;
+
+		this.host.applyConversationUpdate(id, {
+			mcpServerOverrides: overrides
 		});
 
-		this.host.applyConversationUpdate(this.host.activeConversation.id, {
-			mcpServerOverrides: newOverrides.length > 0 ? newOverrides : undefined
+		await DatabaseService.updateConversation(id, {
+			mcpServerOverrides: overrides
 		});
 	}
 
@@ -224,11 +229,13 @@ export class ConversationPreferences {
 			return;
 		}
 
-		this.host.applyConversationUpdate(this.host.activeConversation.id, {
+		const id = this.host.activeConversation.id;
+
+		this.host.applyConversationUpdate(id, {
 			reasoningEffort: effort
 		});
 
-		await DatabaseService.updateConversation(this.host.activeConversation.id, {
+		await DatabaseService.updateConversation(id, {
 			reasoningEffort: effort
 		});
 	}
diff --git a/tools/ui/src/lib/stores/index.ts b/tools/ui/src/lib/stores/index.ts
index db227158f..40ddd10bf 100644
--- a/tools/ui/src/lib/stores/index.ts
+++ b/tools/ui/src/lib/stores/index.ts
@@ -22,6 +22,9 @@ export { chatStore } from './chat/index.svelte';
 
 export { draftMessagesStore } from './chat/drafts.svelte';
 
+// CONVERSATION TABS
+export { tabsStore } from './tabs.svelte';
+
 // CONTEXT STATS (active conversation context window usage)
 export { contextStatsStore } from './chat/context-stats.svelte';
 
@@ -40,6 +43,9 @@ export { modelsStore } from './models/index.svelte';
 // SERVER
 export { serverStore } from './server.svelte';
 
+// UI / LAYOUT
+export { uiStore } from './ui.svelte';
+
 // SETTINGS / UI PREFERENCES
 export { settingsStore } from './settings/index.svelte';
 
diff --git a/tools/ui/src/lib/stores/init.ts b/tools/ui/src/lib/stores/init.ts
index d52c34d0f..37ea87b17 100644
--- a/tools/ui/src/lib/stores/init.ts
+++ b/tools/ui/src/lib/stores/init.ts
@@ -1,21 +1,8 @@
-/**
- * Explicit store initialization, run once and shared by every caller.
- *
- * Order matters: migrations run first because they rename and rewrite
- * localStorage keys, so every store that reads localStorage initializes
- * only after they complete. Constructors and module-level side effects
- * stay empty so import order can no longer change startup behavior.
- *
- * The returned promise resolves once the persisted state is in memory, which
- * route loads await before reading settings: they run ahead of the root layout
- * script. The conversation list loads in the background, awaited by the chat
- * page that renders it.
- */
-
 // direct imports, not via the barrel, to avoid circular deps
 import { conversationsStore } from './conversations/index.svelte';
 import { permissionsStore } from './permissions.svelte';
 import { settingsStore } from './settings/index.svelte';
+import { tabsStore } from './tabs.svelte';
 import { toolsStore } from './tools.svelte';
 import { versionStore } from './version.svelte';
 import { browser } from '$app/environment';
@@ -33,7 +20,12 @@ export function initStores(): Promise<void> {
 		permissionsStore.initialize();
 		toolsStore.initialize();
 		void versionStore.initialize();
-		void conversationsStore.initialize();
+
+		// the full conversation list loads in the background; once it is back,
+		// prune persisted tabs against the conversations that still exist
+		void conversationsStore.initialize().then(() => {
+			tabsStore.init(conversationsStore.conversations.map((c) => c.id));
+		});
 	})();
 
 	return startup;
diff --git a/tools/ui/src/lib/stores/tabs.svelte.ts b/tools/ui/src/lib/stores/tabs.svelte.ts
new file mode 100644
index 000000000..04b476cb6
--- /dev/null
+++ b/tools/ui/src/lib/stores/tabs.svelte.ts
@@ -0,0 +1,154 @@
+/**
+ * tabsStore - Reactive State Store for Browser-Style Conversation Tabs
+ *
+ * Tracks which conversations and the new-chat screen are open as tabs in
+ * the chat layout, in order. Real conversation tabs are `#/chat/<id>`
+ * routes; the new-chat tab is the bare `#/` route, represented here by the
+ * `NEW_CHAT_TAB_ID` sentinel (see {@link NEW_CHAT_TAB_ID}).
+ *
+ * **Architecture & Relationships:**
+ * - **conversationsStore**: owns conversation data; calls `removeTabs()` /
+ *   `close()` when conversations are deleted. This store never imports it,
+ *   so there is no circular dependency - tab names are resolved by the
+ *   ChatTabs component from conversationsStore.
+ * - Tab order persists to localStorage and is pruned against the loaded
+ *   conversation list on init. The new-chat tab is kept across reloads.
+ */
+
+import { browser } from '$app/environment';
+import { goto } from '$app/navigation';
+import { CONVERSATION_TABS_LOCALSTORAGE_KEY, NEW_CHAT_TAB_ID, ROUTES } from '$lib/constants';
+import { RouterService } from '$lib/services/router.service';
+import { untrack } from 'svelte';
+
+class TabsStore {
+	/** Ordered tab ids: conversation ids and the `NEW_CHAT_TAB_ID` sentinel */
+	openTabs = $state<string[]>([]);
+
+	/** False until init() has read the persisted tabs; save() is a no-op before that */
+	private initialized = false;
+
+	/** Navigate to a tab (the new-chat sentinel maps to the bare `#/` route) */
+	async activate(id: string): Promise<void> {
+		await goto(id === NEW_CHAT_TAB_ID ? ROUTES.START : RouterService.chat(id));
+	}
+
+	/** Remove all tabs (e.g. after deleting all conversations) */
+	clear(): void {
+		this.openTabs = [];
+		this.save();
+	}
+
+	/**
+	 * Close a tab. When it belongs to the active route, navigate to the left
+	 * neighbor (or the right one when the closed tab was leftmost), falling
+	 * back to the new-chat screen when no tabs remain.
+	 * @param id - Tab id to close
+	 * @param activeTabId - Tab id of the current route, if any
+	 */
+	async close(id: string, activeTabId: string | null): Promise<void> {
+		const idx = this.openTabs.indexOf(id);
+
+		if (idx === -1) {
+			// tab not tracked (e.g. Conversation tabs are off); still fall back to
+			// the new-chat screen when closing the active conversation
+			if (id === activeTabId) {
+				await goto(ROUTES.START);
+			}
+
+			return;
+		}
+
+		this.openTabs = this.openTabs.filter((tabId) => tabId !== id);
+		this.save();
+
+		if (id !== activeTabId) return;
+
+		const target = (idx > 0 ? this.openTabs[idx - 1] : this.openTabs[0]) ?? null;
+
+		if (target) {
+			await goto(target === NEW_CHAT_TAB_ID ? ROUTES.START : RouterService.chat(target));
+		} else {
+			await goto(ROUTES.START);
+		}
+	}
+
+	/**
+	 * Load persisted tabs, dropping conversation ids that no longer exist.
+	 * Called once from initStores() after conversations are loaded.
+	 * Merges with (rather than replaces) current openTabs: the chat layout
+	 * syncs the route's tab before this async init completes, and replacing
+	 * here would drop it.
+	 * @param validIds - Ids of conversations present in the database
+	 */
+	init(validIds: string[]): void {
+		if (!browser) return;
+
+		// the new-chat sentinel is a pseudo-tab, not a conversation, but it is
+		// still kept so a reload on `#/` does not drop the tab the user is on
+		const isLive = (id: string) => validIds.includes(id) || id === NEW_CHAT_TAB_ID;
+		const persisted = this.load().filter(isLive);
+		// tabs already in openTabs come from the live route, so they stay as they
+		// are: `validIds` is a snapshot and a conversation created while the list
+		// was loading is not in it
+		const extras = this.openTabs.filter((id) => !persisted.includes(id));
+
+		this.openTabs = [...persisted, ...extras];
+		this.initialized = true;
+		this.save();
+	}
+
+	/**
+	 * Remove tabs without navigating. Used when conversations are deleted
+	 * while some other conversation stays open.
+	 * @param ids - Tab ids to drop
+	 */
+	removeTabs(ids: string[]): void {
+		const removed = new Set(ids);
+		const next = this.openTabs.filter((id) => !removed.has(id));
+
+		if (next.length !== this.openTabs.length) {
+			this.openTabs = next;
+			this.save();
+		}
+	}
+
+	/**
+	 * Sync the tab strip with the route. Called from the chat layout on every
+	 * navigation, so any way of reaching a conversation or new-chat tab opens
+	 * a tab for it.
+	 * @param id - The conversation (or temporary new-chat) id of the route
+	 */
+	syncWithRoute(id: string): void {
+		// untrack: callers invoke this from an effect keyed on the route, and
+		// reading openTabs here would subscribe that effect to openTabs too -
+		// closing the active tab would then re-run the effect and re-add the tab
+		untrack(() => {
+			if (!this.openTabs.includes(id)) {
+				this.openTabs = [...this.openTabs, id];
+				this.save();
+			}
+		});
+	}
+
+	private load(): string[] {
+		try {
+			const raw = localStorage.getItem(CONVERSATION_TABS_LOCALSTORAGE_KEY);
+			const parsed: unknown = raw ? JSON.parse(raw) : [];
+
+			return Array.isArray(parsed) ? parsed.filter((id) => typeof id === 'string') : [];
+		} catch {
+			return [];
+		}
+	}
+
+	private save(): void {
+		// never write before init has read the persisted tabs, or an early
+		// route sync (layout effect runs before async init) would clobber them
+		if (!browser || !this.initialized) return;
+
+		localStorage.setItem(CONVERSATION_TABS_LOCALSTORAGE_KEY, JSON.stringify(this.openTabs));
+	}
+}
+
+export const tabsStore = new TabsStore();
diff --git a/tools/ui/src/lib/stores/ui.svelte.ts b/tools/ui/src/lib/stores/ui.svelte.ts
new file mode 100644
index 000000000..57fcfb91e
--- /dev/null
+++ b/tools/ui/src/lib/stores/ui.svelte.ts
@@ -0,0 +1,14 @@
+/**
+ * uiStore - Shared UI/layout state
+ *
+ * Holds cross-component UI state that does not belong to a single component
+ * (e.g. the desktop sidebar's expanded/collapsed state, which the sidebar
+ * controls and the chat tab bar reacts to).
+ */
+
+class UiStore {
+	/** Whether the desktop sidebar is expanded (open). */
+	isSidebarExpanded = $state(false);
+}
+
+export const uiStore = new UiStore();
diff --git a/tools/ui/src/lib/types/navigation.d.ts b/tools/ui/src/lib/types/navigation.d.ts
index 357e060dc..af1e21d0a 100644
--- a/tools/ui/src/lib/types/navigation.d.ts
+++ b/tools/ui/src/lib/types/navigation.d.ts
@@ -1,3 +1,4 @@
+import type { SidebarAction } from '$lib/enums';
 import type { Component } from 'svelte';
 
 /**
@@ -7,6 +8,8 @@ export interface DesktopIconStripItem {
 	icon: Component;
 	tooltip: string;
 	route?: string;
+	/** Custom action handled by the sidebar, e.g. opening a new-chat tab */
+	action?: SidebarAction;
 	activeRouteId?: string;
 	activeRoutePrefix?: string;
 	activeUrlIncludes?: string;
diff --git a/tools/ui/src/routes/(chat)/+layout.svelte b/tools/ui/src/routes/(chat)/+layout.svelte
index 37aa03582..5be56dd56 100644
--- a/tools/ui/src/routes/(chat)/+layout.svelte
+++ b/tools/ui/src/routes/(chat)/+layout.svelte
@@ -1,12 +1,34 @@
 <script lang="ts">
 	import { page } from '$app/state';
-	import { ChatScreen } from '$lib/components/app';
+	import { ChatScreen, ChatTabs } from '$lib/components/app';
+	import { NEW_CHAT_TAB_ID } from '$lib/constants';
+	import { settingsStore, tabsStore } from '$lib/stores';
 
 	let { children } = $props();
 
+	// the new-chat screen is the bare `#/` route (no conversation id)
 	let showCenteredEmpty = $derived(!page.params.id);
+
+	let showTabs = $derived(
+		Boolean(settingsStore.config.conversationTabs) &&
+			(page.params.id || tabsStore.openTabs.some((id) => id !== NEW_CHAT_TAB_ID))
+	);
+
+	$effect(() => {
+		const id = page.params.id ?? (page.route.id === '/(chat)' ? NEW_CHAT_TAB_ID : undefined);
+
+		if (id && settingsStore.config.conversationTabs) {
+			tabsStore.syncWithRoute(id);
+		}
+	});
 </script>
 
-<ChatScreen {showCenteredEmpty} />
+<div class={showTabs ? 'md:[--chat-tabs-offset:1.25rem]' : ''}>
+	{#if showTabs}
+		<ChatTabs />
+	{/if}
+
+	<ChatScreen {showCenteredEmpty} />
+</div>
 
 {@render children?.()}
diff --git a/tools/ui/src/routes/(chat)/+page.svelte b/tools/ui/src/routes/(chat)/+page.svelte
index de8574e35..95a3e1947 100644
--- a/tools/ui/src/routes/(chat)/+page.svelte
+++ b/tools/ui/src/routes/(chat)/+page.svelte
@@ -8,24 +8,19 @@
 
 	let qParam = $derived(page.url.searchParams.get(URL_PARAMS.QUERY));
 	let modelParam = $derived(page.url.searchParams.get(URL_PARAMS.MODEL));
-	let newChatParam = $derived(page.url.searchParams.get(URL_PARAMS.NEW_CHAT));
 	let loadParam = $derived(page.url.searchParams.get(URL_PARAMS.LOAD));
 
-	// Dialog state for model not available error
 	let showModelNotAvailable = $state(false);
 	let requestedModelName = $state('');
 	let availableModelNames = $derived(modelsStore.models.map((m) => m.model));
 
-	/**
-	 * Clear URL params after message is sent to prevent re-sending on refresh
-	 */
+	// Clear params after handling the deep link so a refresh does not replay them
 	function clearUrlParams() {
 		const url = new URL(page.url);
 
 		url.searchParams.delete(URL_PARAMS.QUERY);
 		url.searchParams.delete(URL_PARAMS.MODEL);
 		url.searchParams.delete(URL_PARAMS.LOAD);
-		url.searchParams.delete(URL_PARAMS.NEW_CHAT);
 
 		replaceState(url.toString(), {});
 	}
@@ -40,8 +35,8 @@
 				try {
 					await modelsStore.selectModelById(model.id);
 
-					// with ?load=true, start loading right away so the model is ready sooner;
-					// not awaited, so the UI stays usable during the load
+					// with ?load=true in router mode, start loading right away so the
+					// model is ready sooner; not awaited so the UI stays usable
 					if (
 						loadParam === 'true' &&
 						serverStore.isRouterMode &&
@@ -66,11 +61,12 @@
 			}
 		}
 
-		// Handle ?q= parameter - create new conversation and send message
+		// ?q= creates the conversation, the chat route sends the prompt once the
+		// conversation id is in the URL
 		if (qParam !== null) {
 			await conversationsStore.createConversation();
 			clearUrlParams();
-		} else if (modelParam || newChatParam === 'true') {
+		} else if (modelParam) {
 			clearUrlParams();
 		}
 	}
@@ -85,7 +81,7 @@
 
 		await modelsStore.fetch();
 
-		if (qParam !== null || modelParam !== null || newChatParam === 'true') {
+		if (qParam !== null || modelParam !== null) {
 			await handleUrlParams();
 		}
 
diff --git a/tools/ui/src/routes/+layout.svelte b/tools/ui/src/routes/+layout.svelte
index f87bbe26a..16adfe332 100644
--- a/tools/ui/src/routes/+layout.svelte
+++ b/tools/ui/src/routes/+layout.svelte
@@ -11,6 +11,7 @@
 		FAVICON_PATHS,
 		FAVICON_SELECTORS,
 		HEADERS,
+		NEW_CHAT_TAB_ID,
 		ROUTES,
 		SETTINGS_KEYS,
 		TOOLTIP_DELAY_DURATION
@@ -26,6 +27,7 @@
 		modelsStore,
 		serverStore,
 		settingsStore,
+		tabsStore,
 		versionStore
 	} from '$lib/stores';
 	import { initStores } from '$lib/stores/init';
@@ -74,6 +76,27 @@
 		}
 	}
 
+	function navigateToTab(direction: -1 | 1) {
+		// only makes sense with conversation tabs enabled
+		if (!settingsStore.config.conversationTabs) return;
+
+		const openTabs = tabsStore.openTabs;
+
+		if (openTabs.length === 0) return;
+
+		const activeId = page.params.id ?? NEW_CHAT_TAB_ID;
+		const idx = openTabs.indexOf(activeId);
+		// active tab not in list (e.g. a non-chat route): start from an edge
+		const targetIdx =
+			idx === -1
+				? direction === 1
+					? 0
+					: openTabs.length - 1
+				: (idx + direction + openTabs.length) % openTabs.length;
+
+		void tabsStore.activate(openTabs[targetIdx]);
+	}
+
 	function navigateToConversation(direction: -1 | 1) {
 		const allConvs = conversationsStore.conversations;
 
@@ -96,15 +119,31 @@
 		if (targetIdx >= 0 && targetIdx < allConvs.length) {
 			goto(RouterService.chat(allConvs[targetIdx].id));
 		} else {
-			goto(ROUTES.NEW_CHAT);
+			conversationsStore.openNewChat();
 		}
 	}
 
+	// navigating away from the new-chat screen drops its tab, so it does not
+	// linger once the user moves to a real conversation or another route
+	let previousChatId = $state<string | undefined>(undefined);
+
+	$effect(() => {
+		const id = page.params.id ?? (page.route.id === '/(chat)' ? NEW_CHAT_TAB_ID : undefined);
+		const prev = untrack(() => previousChatId);
+
+		previousChatId = id;
+
+		if (id !== prev && prev && settingsStore.config.conversationTabs && prev === NEW_CHAT_TAB_ID) {
+			untrack(() => tabsStore.removeTabs([NEW_CHAT_TAB_ID]));
+		}
+	});
 	// Global keyboard shortcuts
 	const { handleKeydown } = useKeyboardShortcuts({
 		editActiveConversation: () => chatSidebar?.editActiveConversation?.(),
 		navigateToNextConversation: () => navigateToConversation(1),
-		navigateToPrevConversation: () => navigateToConversation(-1)
+		navigateToNextTab: () => navigateToTab(1),
+		navigateToPrevConversation: () => navigateToConversation(-1),
+		navigateToPrevTab: () => navigateToTab(-1)
 	});
 
 	function checkApiKey() {
diff --git a/tools/ui/src/routes/search/+page.svelte b/tools/ui/src/routes/search/+page.svelte
index a882e5552..548b9911a 100644
--- a/tools/ui/src/routes/search/+page.svelte
+++ b/tools/ui/src/routes/search/+page.svelte
@@ -21,10 +21,10 @@
 	});
 
 	// Search page is intended for mobile; on desktop the sidebar already exposes
-	// in-place search, so bounce back to a chat.
+	// in-place search, so bounce back to a new-chat tab without a history entry.
 	$effect(() => {
 		if (browser && !deviceStore.isMobile) {
-			goto(ROUTES.NEW_CHAT, { replaceState: true });
+			goto(ROUTES.START, { replaceState: true });
 		}
 	});
 
@@ -66,7 +66,7 @@
 		if (history.length > 1) {
 			history.back();
 		} else {
-			goto(ROUTES.NEW_CHAT);
+			conversationsStore.openNewChat();
 		}
 	}
 </script>
diff --git a/tools/ui/tests/stories/a11y/HorizontalScrollCarousel.a11y.stories.svelte b/tools/ui/tests/stories/a11y/ScrollCarousel.a11y.stories.svelte
similarity index 80%
rename from tools/ui/tests/stories/a11y/HorizontalScrollCarousel.a11y.stories.svelte
rename to tools/ui/tests/stories/a11y/ScrollCarousel.a11y.stories.svelte
index ef5abeafa..b9bf5afbc 100644
--- a/tools/ui/tests/stories/a11y/HorizontalScrollCarousel.a11y.stories.svelte
+++ b/tools/ui/tests/stories/a11y/ScrollCarousel.a11y.stories.svelte
@@ -1,15 +1,16 @@
 <script module lang="ts">
 	import { defineMeta } from '@storybook/addon-svelte-csf';
-	import HorizontalScrollCarousel from '$lib/components/app/misc/HorizontalScrollCarousel.svelte';
+	import { ScrollCarousel } from '$lib/components/app';
+	import { ScrollCarouselVariant } from '$lib/enums';
 	import { expect, waitFor } from 'storybook/test';
 
 	const { Story } = defineMeta({
-		component: HorizontalScrollCarousel,
+		component: ScrollCarousel,
 		parameters: {
 			layout: 'centered'
 		},
 		tags: ['!dev'],
-		title: 'Components/HorizontalScrollCarousel/Accessibility'
+		title: 'Components/ScrollCarousel/Accessibility'
 	});
 </script>
 
@@ -33,10 +34,10 @@
 >
 	<div>
 		<button type="button">before</button>
-		<HorizontalScrollCarousel class="w-96">
+		<ScrollCarousel class="w-96" variant={ScrollCarouselVariant.CENTER}>
 			<div class="h-12 w-12 shrink-0 bg-muted"></div>
 			<div class="h-12 w-12 shrink-0 bg-muted"></div>
-		</HorizontalScrollCarousel>
+		</ScrollCarousel>
 		<button type="button">after</button>
 	</div>
 </Story>
@@ -60,10 +61,10 @@
 >
 	<div>
 		<button type="button">before</button>
-		<HorizontalScrollCarousel class="w-48">
+		<ScrollCarousel class="w-48" variant={ScrollCarouselVariant.CENTER}>
 			{#each [...Array(20).keys()] as i (i)}
 				<div class="h-12 w-24 shrink-0 bg-muted">{i}</div>
 			{/each}
-		</HorizontalScrollCarousel>
+		</ScrollCarousel>
 	</div>
 </Story>
diff --git a/vendor/sheredom/subprocess.h b/vendor/sheredom/subprocess.h
index c3af8a498..67420c103 100644
--- a/vendor/sheredom/subprocess.h
+++ b/vendor/sheredom/subprocess.h
@@ -275,13 +275,46 @@ subprocess_weak int subprocess_alive(struct subprocess_s *const process);
 #include <unistd.h>
 #endif
 
+#if defined(__NetBSD__)
+#include <sys/param.h>
+#endif
+
+/* Which spelling of the chdir file action the platform provides, if any.
+   POSIX 2024 standardised posix_spawn_file_actions_addchdir; implementations
+   that shipped it earlier called it ..._np. macOS 26 and NetBSD 10 use the
+   standard name, glibc 2.29+, macOS 10.15+ and FreeBSD 13.1+ use the _np name,
+   and AIX, NetBSD 9 and older, and OpenBSD provide neither. */
+#if !defined(SUBPROCESS_ADDCHDIR_IS_POSIX)
+#if (defined(__APPLE__) && MAC_OS_X_VERSION_MIN_REQUIRED >= 260000) ||        \
+    (defined(__NetBSD__) && __NetBSD_Version__ >= 1000000000)
+#define SUBPROCESS_ADDCHDIR_IS_POSIX 1
+#else
+#define SUBPROCESS_ADDCHDIR_IS_POSIX 0
+#endif
+#endif
+
+/* Whether to launch the child with fork()+exec() instead of posix_spawn(),
+   for platforms with no posix_spawn_file_actions_addchdir under either
+   spelling: the child chdir()s before exec, and a close-on-exec pipe carries
+   exec's errno back. Define this yourself to force either implementation. */
+#if !defined(SUBPROCESS_SPAWN_VIA_FORK)
+#if defined(_AIX) || defined(__OpenBSD__) ||                                  \
+    (defined(__NetBSD__) && (__NetBSD_Version__ < 1000000000))
+#define SUBPROCESS_SPAWN_VIA_FORK 1
+#else
+#define SUBPROCESS_SPAWN_VIA_FORK 0
+#endif
+#endif
+
 /* Whether subprocess_create_ex can honour process_cwd. glibc only gained
    posix_spawn_file_actions_addchdir_np in 2.29, and macOS in 10.15; the SDKs
    mark it unavailable on iOS, tvOS and watchOS, where the undefined version
    macro folds to 0 and so answers correctly. Define this yourself to override
    the detection, for instance on musl older than 1.1.24. */
 #if !defined(SUBPROCESS_HAVE_CWD)
-#if defined(__GLIBC__)
+#if SUBPROCESS_SPAWN_VIA_FORK
+#define SUBPROCESS_HAVE_CWD 1
+#elif defined(__GLIBC__)
 #if __GLIBC_PREREQ(2, 29)
 #define SUBPROCESS_HAVE_CWD 1
 #else
@@ -294,10 +327,13 @@ subprocess_weak int subprocess_alive(struct subprocess_s *const process);
 #endif
 #endif
 
-/* Whether posix_spawn reports a failed exec back to the caller. glibc only
-   started doing so in 2.24; before that the child silently exits with 127. */
+/* Whether a failed exec is reported back to the caller. The fork() path always
+   reports it through its error pipe. glibc's posix_spawn only started doing so
+   in 2.24; before that the child silently exits with 127. */
 #if !defined(SUBPROCESS_SPAWN_REPORTS_EXEC_ERRORS)
-#if defined(__GLIBC__)
+#if SUBPROCESS_SPAWN_VIA_FORK
+#define SUBPROCESS_SPAWN_REPORTS_EXEC_ERRORS 1
+#elif defined(__GLIBC__)
 #if __GLIBC_PREREQ(2, 24)
 #define SUBPROCESS_SPAWN_REPORTS_EXEC_ERRORS 1
 #else
@@ -342,6 +378,14 @@ typedef intptr_t subprocess_intptr_t;
 typedef size_t subprocess_size_t;
 #endif
 
+/* SIZE_T is ULONG_PTR, which is not size_t: on Win32 both are 32 bits wide but
+   unsigned long and unsigned int are still distinct types. */
+#ifdef _WIN64
+typedef subprocess_size_t subprocess_ulongptr_t;
+#else
+typedef unsigned long subprocess_ulongptr_t;
+#endif
+
 #ifdef __clang__
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Wreserved-identifier"
@@ -351,6 +395,7 @@ typedef struct _PROCESS_INFORMATION *LPPROCESS_INFORMATION;
 typedef struct _SECURITY_ATTRIBUTES *LPSECURITY_ATTRIBUTES;
 typedef struct _STARTUPINFOW *LPSTARTUPINFOW;
 typedef struct _OVERLAPPED *LPOVERLAPPED;
+typedef struct _PROC_THREAD_ATTRIBUTE_LIST *LPPROC_THREAD_ATTRIBUTE_LIST;
 
 #ifdef __clang__
 #pragma clang diagnostic pop
@@ -402,6 +447,11 @@ struct subprocess_startup_info_s {
   void *hStdError;
 };
 
+struct subprocess_startup_info_ex_s {
+  struct subprocess_startup_info_s startupInfo;
+  void *attributeList;
+};
+
 struct subprocess_overlapped_s {
   uintptr_t Internal;
   uintptr_t InternalHigh;
@@ -451,6 +501,14 @@ __declspec(dllimport) int __stdcall CreateProcessW(
     const subprocess_wchar_t *, subprocess_wchar_t *, LPSECURITY_ATTRIBUTES,
     LPSECURITY_ATTRIBUTES, int, unsigned long, void *,
     const subprocess_wchar_t *, LPSTARTUPINFOW, LPPROCESS_INFORMATION);
+__declspec(dllimport) int __stdcall
+InitializeProcThreadAttributeList(LPPROC_THREAD_ATTRIBUTE_LIST, unsigned long,
+                                  unsigned long, subprocess_ulongptr_t *);
+__declspec(dllimport) int __stdcall UpdateProcThreadAttribute(
+    LPPROC_THREAD_ATTRIBUTE_LIST, unsigned long, subprocess_ulongptr_t, void *,
+    subprocess_ulongptr_t, void *, subprocess_ulongptr_t *);
+__declspec(dllimport) void __stdcall
+DeleteProcThreadAttributeList(LPPROC_THREAD_ATTRIBUTE_LIST);
 __declspec(dllimport) int __stdcall MultiByteToWideChar(
     unsigned int, unsigned long, const char *, int, subprocess_wchar_t *, int);
 __declspec(dllimport) int __stdcall CloseHandle(void *);
@@ -667,12 +725,104 @@ int subprocess_create_named_pipe_helper(void **rd, void **wr) {
 }
 #endif
 
+#if !defined(_WIN32)
+/* Move a pipe end off 0, 1 or 2. Duplicating a descriptor onto itself is a
+   no-op, so a pipe end already sitting on a standard descriptor would keep its
+   FD_CLOEXEC and be closed by exec, leaving the child without that stream. */
+static int subprocess_fds_above_std(int fds[2]) {
+  int fd_flags;
+  int index;
+  int moved;
+  int saved_errno;
+
+  for (index = 0; index < 2; index++) {
+    if (fds[index] > STDERR_FILENO) {
+      continue;
+    }
+
+    moved = fcntl(fds[index], F_DUPFD, STDERR_FILENO + 1);
+    if (-1 != moved) {
+      fd_flags = fcntl(moved, F_GETFD, 0);
+      if ((-1 == fd_flags) ||
+          (-1 == fcntl(moved, F_SETFD, fd_flags | FD_CLOEXEC))) {
+        saved_errno = errno;
+        close(moved);
+        errno = saved_errno;
+        moved = -1;
+      }
+    }
+
+    if (-1 == moved) {
+      saved_errno = errno;
+      close(fds[0]);
+      close(fds[1]);
+      fds[0] = -1;
+      fds[1] = -1;
+      errno = saved_errno;
+      return -1;
+    }
+
+    close(fds[index]);
+    fds[index] = moved;
+  }
+
+  return 0;
+}
+
+/* Create pipes with close-on-exec set so later subprocesses do not inherit
+   descriptors belonging to subprocesses which are already running. */
+static int subprocess_pipe_cloexec(int fds[2]) {
+  int fd_flags;
+  int index;
+  int saved_errno;
+
+#if defined(__linux__) || defined(__FreeBSD__) || defined(__NetBSD__) ||       \
+    defined(__OpenBSD__) || defined(__DragonFly__) ||                         \
+    (defined(__sun) && defined(__SVR4))
+  if (0 == pipe2(fds, O_CLOEXEC)) {
+    return subprocess_fds_above_std(fds);
+  }
+
+  /* Older kernels can lack pipe2 even when the C library declares it. */
+  if (ENOSYS != errno) {
+    return -1;
+  }
+#endif
+
+  if (0 != pipe(fds)) {
+    return -1;
+  }
+
+  for (index = 0; index < 2; index++) {
+    fd_flags = fcntl(fds[index], F_GETFD, 0);
+    if ((-1 == fd_flags) ||
+        (-1 == fcntl(fds[index], F_SETFD, fd_flags | FD_CLOEXEC))) {
+      saved_errno = errno;
+      close(fds[0]);
+      close(fds[1]);
+      fds[0] = -1;
+      fds[1] = -1;
+      errno = saved_errno;
+      return -1;
+    }
+  }
+
+  return subprocess_fds_above_std(fds);
+}
+#endif
+
 int subprocess_create(const char *const commandLine[], int options,
                       struct subprocess_s *const out_process) {
   return subprocess_create_ex(commandLine, options, SUBPROCESS_NULL,
                               SUBPROCESS_NULL, out_process);
 }
 
+#if SUBPROCESS_SPAWN_VIA_FORK
+/* Not every platform declares execvpe: AIX exports it from libc without ever
+   naming it in a header, and glibc hides it behind _GNU_SOURCE. */
+extern int execvpe(const char *, char *const *, char *const *);
+#endif
+
 int subprocess_create_ex(const char *const commandLine[], int options,
                          const char *const environment[],
                          const char *const process_cwd,
@@ -692,6 +842,7 @@ int subprocess_create_ex(const char *const commandLine[], int options,
   subprocess_size_t bs_run;
   unsigned long flags = 0;
   unsigned long last_error = 0;
+  int attribute_list_initialized = 0;
   int result = subprocess_error_unknown;
   const unsigned int codePageUtf8 = 65001;
   const unsigned long mbErrInvalidChars = 0x00000008;
@@ -699,6 +850,8 @@ int subprocess_create_ex(const char *const commandLine[], int options,
   const unsigned long handleFlagInherit = 0x00000001;
   const unsigned long createNoWindow = 0x08000000;
   const unsigned long createUnicodeEnvironment = 0x00000400;
+  const unsigned long extendedStartupInfoPresent = 0x00080000;
+  const subprocess_size_t procThreadAttributeHandleList = 0x00020002;
   struct subprocess_subprocess_information_s processInfo = {SUBPROCESS_NULL,
                                                             SUBPROCESS_NULL, 0,
                                                             0};
@@ -706,6 +859,11 @@ int subprocess_create_ex(const char *const commandLine[], int options,
                                                     SUBPROCESS_NULL, 1};
   subprocess_wchar_t empty_environment[2] = {0, 0};
   subprocess_wchar_t *used_environment = SUBPROCESS_NULL;
+  subprocess_ulongptr_t attribute_list_size = 0;
+  subprocess_size_t inherited_handle_count = 0;
+  LPPROC_THREAD_ATTRIBUTE_LIST attribute_list = SUBPROCESS_NULL;
+  void *inherited_handles[3];
+  struct subprocess_startup_info_ex_s startInfoEx;
   struct subprocess_startup_info_s startInfo = {0,
                                                 SUBPROCESS_NULL,
                                                 SUBPROCESS_NULL,
@@ -1080,6 +1238,44 @@ int subprocess_create_ex(const char *const commandLine[], int options,
     }
   }
 
+  /* Restrict inheritance to this subprocess's standard streams. Without a
+     handle list, concurrent subprocess_create calls can inherit each other's
+     temporarily-inheritable child pipe handles. */
+  inherited_handles[inherited_handle_count++] = startInfo.hStdInput;
+  inherited_handles[inherited_handle_count++] = startInfo.hStdOutput;
+  if (startInfo.hStdError != startInfo.hStdOutput) {
+    inherited_handles[inherited_handle_count++] = startInfo.hStdError;
+  }
+
+  InitializeProcThreadAttributeList(SUBPROCESS_NULL, 1, 0,
+                                    &attribute_list_size);
+  if (0 == attribute_list_size) {
+    result = subprocess_error_spawn;
+    goto cleanup;
+  }
+
+  attribute_list = SUBPROCESS_PTR_CAST(LPPROC_THREAD_ATTRIBUTE_LIST,
+                                       _alloca(attribute_list_size));
+  if (!attribute_list || !InitializeProcThreadAttributeList(
+                             attribute_list, 1, 0, &attribute_list_size)) {
+    result = subprocess_error_spawn;
+    goto cleanup;
+  }
+  attribute_list_initialized = 1;
+
+  if (!UpdateProcThreadAttribute(
+          attribute_list, 0, procThreadAttributeHandleList, inherited_handles,
+          inherited_handle_count * sizeof(inherited_handles[0]),
+          SUBPROCESS_NULL, SUBPROCESS_NULL)) {
+    result = subprocess_error_spawn;
+    goto cleanup;
+  }
+
+  startInfoEx.startupInfo = startInfo;
+  startInfoEx.startupInfo.cb = sizeof(startInfoEx);
+  startInfoEx.attributeList = attribute_list;
+  flags |= extendedStartupInfoPresent;
+
   if (!CreateProcessW(
           SUBPROCESS_NULL,
           commandLineCombinedWide, // command line
@@ -1090,7 +1286,7 @@ int subprocess_create_ex(const char *const commandLine[], int options,
           used_environment,        // used environment
           process_cwd_wide,        // use specified current directory
           SUBPROCESS_PTR_CAST(LPSTARTUPINFOW,
-                              &startInfo), // STARTUPINFO pointer
+                              &startInfoEx), // STARTUPINFOEX pointer
           SUBPROCESS_PTR_CAST(LPPROCESS_INFORMATION, &processInfo))) {
     result = subprocess_error_from_windows_error(GetLastError());
     if (subprocess_error_unknown == result) {
@@ -1099,6 +1295,9 @@ int subprocess_create_ex(const char *const commandLine[], int options,
     goto cleanup;
   }
 
+  DeleteProcThreadAttributeList(attribute_list);
+  attribute_list_initialized = 0;
+
   out_process->hProcess = processInfo.hProcess;
   processInfo.hProcess = SUBPROCESS_NULL;
 
@@ -1128,6 +1327,10 @@ int subprocess_create_ex(const char *const commandLine[], int options,
 cleanup:
   last_error = GetLastError();
 
+  if (attribute_list_initialized) {
+    DeleteProcThreadAttributeList(attribute_list);
+  }
+
   if (subprocess_error_unknown == result) {
     result = subprocess_error_from_windows_error(last_error);
   }
@@ -1173,15 +1376,20 @@ cleanup:
   int stderrfd[2] = {-1, -1};
   int fd, fd_flags;
   int async_no_wait;
-  int actions_created = 0;
   int result = subprocess_error_unknown;
   int saved_errno = 0;
-  int posix_error;
   pid_t child = 0;
   extern char **environ;
   char *const empty_environment[1] = {SUBPROCESS_NULL};
-  posix_spawn_file_actions_t actions;
   char *const *used_environment;
+#if SUBPROCESS_SPAWN_VIA_FORK
+  /* Pipe used to relay the child's exec() errno back to the parent. */
+  int exec_errfd[2] = {-1, -1};
+#else
+  int actions_created = 0;
+  int posix_error;
+  posix_spawn_file_actions_t actions;
+#endif
 
   async_no_wait = subprocess_option_enable_async_no_wait ==
                   (options & subprocess_option_enable_async_no_wait);
@@ -1202,13 +1410,13 @@ cleanup:
 
   memset(out_process, 0, sizeof(*out_process));
 
-  if (0 != pipe(stdinfd)) {
+  if (0 != subprocess_pipe_cloexec(stdinfd)) {
     saved_errno = errno;
     result = subprocess_error_pipe;
     goto cleanup;
   }
 
-  if (0 != pipe(stdoutfd)) {
+  if (0 != subprocess_pipe_cloexec(stdoutfd)) {
     saved_errno = errno;
     result = subprocess_error_pipe;
     goto cleanup;
@@ -1216,7 +1424,7 @@ cleanup:
 
   if (subprocess_option_combined_stdout_stderr !=
       (options & subprocess_option_combined_stdout_stderr)) {
-    if (0 != pipe(stderrfd)) {
+    if (0 != subprocess_pipe_cloexec(stderrfd)) {
       saved_errno = errno;
       result = subprocess_error_pipe;
       goto cleanup;
@@ -1240,6 +1448,136 @@ cleanup:
     used_environment = empty_environment;
   }
 
+#if SUBPROCESS_SPAWN_VIA_FORK
+  /* fork()+exec() instead of posix_spawn, so the child can chdir() first.
+     exec_errfd[1] is close-on-exec: a successful exec closes it and the parent
+     reads EOF; a failed exec writes errno through it before _exit. */
+  if (0 != pipe(exec_errfd)) {
+    saved_errno = errno;
+    result = subprocess_error_pipe;
+    goto cleanup;
+  }
+
+  if (-1 == fcntl(exec_errfd[1], F_SETFD, FD_CLOEXEC)) {
+    saved_errno = errno;
+    result = subprocess_error_spawn;
+    goto cleanup;
+  }
+
+  child = fork();
+
+  if (child < 0) {
+    saved_errno = errno;
+    result = subprocess_error_spawn;
+    goto cleanup;
+  }
+
+  if (0 == child) {
+    /* Child. Everything below must stay async-signal-safe: after fork() in a
+       threaded process only such functions may be called before exec. */
+    int child_errno;
+
+    close(exec_errfd[0]);
+
+    if ((-1 == dup2(stdinfd[0], STDIN_FILENO)) ||
+        (-1 == dup2(stdoutfd[1], STDOUT_FILENO))) {
+      goto child_failed;
+    }
+
+    if (subprocess_option_combined_stdout_stderr ==
+        (options & subprocess_option_combined_stdout_stderr)) {
+      if (-1 == dup2(STDOUT_FILENO, STDERR_FILENO)) {
+        goto child_failed;
+      }
+    } else {
+      if (-1 == dup2(stderrfd[1], STDERR_FILENO)) {
+        goto child_failed;
+      }
+    }
+
+    /* The originals are only closed once they have been duplicated, so that a
+       pipe end that already sits on 0, 1 or 2 is not closed out from under us. */
+    if (stdinfd[0] > STDERR_FILENO) {
+      close(stdinfd[0]);
+    }
+    if (stdinfd[1] > STDERR_FILENO) {
+      close(stdinfd[1]);
+    }
+    if (stdoutfd[0] > STDERR_FILENO) {
+      close(stdoutfd[0]);
+    }
+    if (stdoutfd[1] > STDERR_FILENO) {
+      close(stdoutfd[1]);
+    }
+    if (stderrfd[0] > STDERR_FILENO) {
+      close(stderrfd[0]);
+    }
+    if (stderrfd[1] > STDERR_FILENO) {
+      close(stderrfd[1]);
+    }
+
+    if (process_cwd && (0 != chdir(process_cwd))) {
+      goto child_failed;
+    }
+
+#ifdef __clang__
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wcast-qual"
+#pragma clang diagnostic ignored "-Wold-style-cast"
+#endif
+    if (subprocess_option_search_user_path ==
+        (options & subprocess_option_search_user_path)) {
+      execvpe(commandLine[0],
+              SUBPROCESS_CONST_CAST(char *const *, commandLine),
+              SUBPROCESS_CONST_CAST(char *const *, used_environment));
+    } else {
+      execve(commandLine[0],
+             SUBPROCESS_CONST_CAST(char *const *, commandLine),
+             SUBPROCESS_CONST_CAST(char *const *, used_environment));
+    }
+#ifdef __clang__
+#pragma clang diagnostic pop
+#endif
+
+  child_failed:
+    child_errno = errno;
+    /* Nothing useful can be done if this write fails; the parent then sees EOF
+       and reports success, exactly as posix_spawn would without exec reporting. */
+    (void)!write(exec_errfd[1], &child_errno, sizeof(child_errno));
+    /* 127 is what POSIX requires posix_spawn's child to exit with when exec
+       fails, so both implementations look the same to a caller. */
+    _exit(127);
+  }
+
+  /* Parent. */
+  close(exec_errfd[1]);
+  exec_errfd[1] = -1;
+
+  {
+    int child_errno = 0;
+    ssize_t bytes_read;
+
+    do {
+      bytes_read = read(exec_errfd[0], &child_errno, sizeof(child_errno));
+    } while ((-1 == bytes_read) && (EINTR == errno));
+
+    close(exec_errfd[0]);
+    exec_errfd[0] = -1;
+
+    if (bytes_read == (ssize_t)sizeof(child_errno)) {
+      /* exec failed in the child. Reap it and surface the reason. */
+      while ((-1 == waitpid(child, SUBPROCESS_NULL, 0)) && (EINTR == errno)) {
+      }
+      child = 0;
+      saved_errno = child_errno;
+      result = subprocess_error_from_errno(child_errno);
+      if (subprocess_error_unknown == result) {
+        result = subprocess_error_spawn;
+      }
+      goto cleanup;
+    }
+  }
+#else
   posix_error = posix_spawn_file_actions_init(&actions);
   if (0 != posix_error) {
     saved_errno = posix_error;
@@ -1253,7 +1591,7 @@ cleanup:
 
   // Set working directory
   if (process_cwd) {
-#if defined(__NetBSD__) || (defined(__APPLE__) && MAC_OS_X_VERSION_MIN_REQUIRED >= 260000)
+#if SUBPROCESS_ADDCHDIR_IS_POSIX
     posix_error = posix_spawn_file_actions_addchdir(&actions, process_cwd);
 #elif !SUBPROCESS_HAVE_CWD
     posix_error = ENOSYS;
@@ -1406,6 +1744,7 @@ cleanup:
 #ifdef __clang__
 #pragma clang diagnostic pop
 #endif
+#endif /* SUBPROCESS_SPAWN_VIA_FORK */
 
   // Close the stdin read end
   close(stdinfd[0]);
@@ -1480,9 +1819,21 @@ cleanup:
     result = subprocess_error_from_errno(saved_errno);
   }
 
+#if SUBPROCESS_SPAWN_VIA_FORK
+  if (-1 != exec_errfd[0]) {
+    close(exec_errfd[0]);
+    exec_errfd[0] = -1;
+  }
+
+  if (-1 != exec_errfd[1]) {
+    close(exec_errfd[1]);
+    exec_errfd[1] = -1;
+  }
+#else
   if (actions_created) {
     posix_spawn_file_actions_destroy(&actions);
   }
+#endif
 
   if (0 != result) {
     if (child) {