diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index d559d70631..0ca9f97151 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -4936,7 +4936,7 @@ class Qwen3VLVisionModel(MmprojModel):
             return
 
         if name.startswith("visual."):
-            return [(self.map_tensor_name(name), data_torch)]
+            yield (self.map_tensor_name(name), data_torch)
         return [] # skip other tensors
 
 
@@ -4975,7 +4975,6 @@ class Qwen3OmniMmprojModel(Qwen3VLVisionModel, Qwen25AudioModel):
                 # transform conv2d bias [n_embd] --> [1, 1, n_embd]
                 data_torch = data_torch.unsqueeze(-1).unsqueeze(-1)
             yield from Qwen25AudioModel.modify_tensors(self, data_torch, name, bid)
-        return []
 
 
 @ModelBase.register("Glm4vForConditionalGeneration", "Glm4vMoeForConditionalGeneration", "GlmOcrForConditionalGeneration")
@@ -5028,14 +5027,20 @@ class Qwen3VLTextModel(Qwen3Model):
 class Qwen3VLMoeTextModel(Qwen3MoeModel):
     model_arch = gguf.MODEL_ARCH.QWEN3VLMOE
 
+    def set_vocab(self):
+        super().set_vocab()
+        # correct BOS/EOS tokens
+        with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f:
+            tokenizer_config = json.load(f)
+            added_tokens = tokenizer_config.get("added_tokens_decoder", {})
+            for token_id, data in added_tokens.items():
+                if data.get("content") == "<|im_end|>":
+                    self.gguf_writer.add_bos_token_id(int(token_id))
+                    self.gguf_writer.add_eos_token_id(int(token_id))
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        if "thinker_config" in self.hparams:
-            vision_config = self.hparams["thinker_config"].get("vision_config", {})
-        else:
-            vision_config = self.hparams.get("vision_config", {})
-        deepstack_layer_num = len(vision_config.get("deepstack_visual_indexes", []))
-        self.gguf_writer.add_num_deepstack_layers(deepstack_layer_num)
+        self.gguf_writer.add_num_deepstack_layers(0)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # Skip vision tensors - they go in the mmproj file
diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index 3178ea6601..12a298e9a8 100644
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -2669,7 +2669,12 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
             } break;
         case PROJECTOR_TYPE_QWEN3A:
             {
-                return 375; // TODO: calculate this
+                // 3x stride-2 conv2d: each step is floor((n-1)/2)+1
+                int n = img->nx;
+                n = (n - 1) / 2 + 1;
+                n = (n - 1) / 2 + 1;
+                n = (n - 1) / 2 + 1;
+                n_patches = n;
             } break;
         case PROJECTOR_TYPE_GLMA:
             {
@@ -3256,7 +3261,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         case PROJECTOR_TYPE_QWEN2A:
             return ctx->model.mm_fc_w->ne[1];
         case PROJECTOR_TYPE_QWEN3A:
-            return ctx->model.mm_2_w->ne[1] * 4; // 4 for deepstack, TODO: do NOT hardcode
+            return ctx->model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_GLMA:
         case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_KIMIVL: