diff --git a/tests/test-llama-archs.cpp b/tests/test-llama-archs.cpp
index ece05ecc3e..d03ba2d6ad 100644
--- a/tests/test-llama-archs.cpp
+++ b/tests/test-llama-archs.cpp
@@ -435,6 +435,13 @@ static int test_backends(const llm_arch target_arch, const size_t seed, const gg
             continue; // TODO tensor shapes
         }
 
+        // FIXME some models are segfaulting with WebGPU:
+#ifdef GGML_USE_WEBGPU
+        if (arch == LLM_ARCH_QWEN3NEXT || arch == LLM_ARCH_QWEN35 || arch == LLM_ARCH_QWEN35MOE || arch == LLM_ARCH_KIMI_LINEAR) {
+            continue;
+        }
+#endif // GGML_USE_WEBGPU
+
         const bool encode = arch == LLM_ARCH_T5;
         for (bool moe : {false, true}) {
             if (moe && !moe_implemented(arch)) {
@@ -454,7 +461,7 @@ static int test_backends(const llm_arch target_arch, const size_t seed, const gg
                 auto model_and_ctx_dev = get_model_and_ctx(gguf_ctx.get(), seed, {dev});
                 const std::vector<float> logits_dev = get_logits(model_and_ctx_dev.first.get(), model_and_ctx_dev.second.get(), tokens, encode);
                 const double nmse_val = nmse(logits_cpu, logits_dev);
-                const bool ok = nmse_val <= 1e-5;
+                const bool ok = nmse_val <= 1e-4;
                 all_ok = all_ok && ok;
                 char nmse_str[10];
                 snprintf(nmse_str, sizeof(nmse_str), "%.2e", nmse_val);
diff --git a/tools/results/results.cpp b/tools/results/results.cpp
index ca69bf305e..e5c5df128e 100644
--- a/tools/results/results.cpp
+++ b/tools/results/results.cpp
@@ -130,7 +130,7 @@ int main(int argc, char ** argv) {
         const double nmse_val = nmse(logits_disk, logits_calc);
         LOG_INF("%s: NMSE=%.3e\n", __func__, nmse_val);
 
-        if (nmse_val > 1e-5) {
+        if (nmse_val > 1e-6) {
             printf("\033[1;31mFAIL\033[0m\n");
             return 1;
         }