diff --git a/gpttype_adapter.cpp b/gpttype_adapter.cpp
index 616dcac8c..ffa70be23 100644
--- a/gpttype_adapter.cpp
+++ b/gpttype_adapter.cpp
@@ -3672,6 +3672,11 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
                     img_start = "<|begin_of_image|>";
                     img_end = "<|end_of_image|>\n\n";
                 }
+                else if(ptype==PROJECTOR_TYPE_GEMMA4V)
+                {
+                    img_start = "<|image>";
+                    img_end = "<image|>\n\n";
+                }
             }
             TokenizeString(img_start, lv.chunk_start_seq, file_format, false);
             TokenizeString(img_end, lv.chunk_end_seq, file_format, false);
diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index 52ef5e79d..0a5e1d5c6 100644
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -4041,6 +4041,10 @@ bool clip_is_llava(const struct clip_ctx * ctx) {
     return ctx->model.hparams.has_llava_projector;
 }
 
+bool clip_is_gemma4(const struct clip_ctx * ctx) { //for kcpp use
+    return ctx->proj_type() == PROJECTOR_TYPE_GEMMA4V;
+}
+
 bool clip_has_vision_encoder(const struct clip_ctx * ctx) {
     return ctx->model.modality == CLIP_MODALITY_VISION;
 }
diff --git a/tools/mtmd/clip.h b/tools/mtmd/clip.h
index 84cdbebc0..7440152a3 100644
--- a/tools/mtmd/clip.h
+++ b/tools/mtmd/clip.h
@@ -112,6 +112,7 @@ bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct
 
 int clip_is_minicpmv(const struct clip_ctx * ctx);
 bool clip_is_glm(const struct clip_ctx * ctx);
+bool clip_is_gemma4(const struct clip_ctx * ctx);
 bool clip_is_mrope(const struct clip_ctx * ctx);
 bool clip_is_llava(const struct clip_ctx * ctx);
 void set_clip_uses_gpu(bool usegpu);
diff --git a/tools/mtmd/llava.cpp b/tools/mtmd/llava.cpp
index e05f295f3..e30cb2465 100644
--- a/tools/mtmd/llava.cpp
+++ b/tools/mtmd/llava.cpp
@@ -78,7 +78,17 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
             max_ny = std::max(max_ny,b);
         }
         image_embd = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, max_nx, max_ny));
-    } else {
+    }
+    else if(clip_is_gemma4(ctx_clip)) //gemma4 vision
+    {
+        const size_t n_imgs = clip_image_f32_batch_n_images(preprocessed_img.get());
+        clip_image_f32 * img_res = clip_image_f32_get_img(preprocessed_img.get(), 0);
+        int n_img_pos = clip_n_output_tokens(ctx_clip, img_res);
+        int nbytes = n_img_pos * clip_n_mmproj_embd(ctx_clip) * sizeof(float);
+        image_embd = (float *)malloc(nbytes);
+    }
+    else
+    {
         image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
     }
     if (!image_embd) {