diff --git a/vllm/model_executor/models/aria.py b/vllm/model_executor/models/aria.py
index 3d07e6b612ca..c6d7f19cbe90 100644
--- a/vllm/model_executor/models/aria.py
+++ b/vllm/model_executor/models/aria.py
@@ -499,8 +499,6 @@ class AriaForConditionalGeneration(nn.Module, SupportsMultiModal):
     model to perform tasks that involve both image and text inputs.
     """
 
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             # mapping for new names in checkpoint saved after transformers v4.52
diff --git a/vllm/model_executor/models/aya_vision.py b/vllm/model_executor/models/aya_vision.py
index 0ada2ed5028b..ee9e210a3240 100644
--- a/vllm/model_executor/models/aya_vision.py
+++ b/vllm/model_executor/models/aya_vision.py
@@ -318,8 +318,6 @@ def _get_layer_index(feature_layer_index: int, num_hidden_layers: int) -> int:
     dummy_inputs=AyaVisionDummyInputsBuilder,
 )
 class AyaVisionForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             # mapping for new names in checkpoint saved after transformers v4.52
diff --git a/vllm/model_executor/models/blip2.py b/vllm/model_executor/models/blip2.py
index f71b9c01d359..1244f97a1bd6 100644
--- a/vllm/model_executor/models/blip2.py
+++ b/vllm/model_executor/models/blip2.py
@@ -523,8 +523,6 @@ def _get_prompt_updates(
 class Blip2ForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsPP, SupportsQuant
 ):
-    merge_by_field_config = True
-
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):
diff --git a/vllm/model_executor/models/chameleon.py b/vllm/model_executor/models/chameleon.py
index 3aa01bb1905f..dfc05a366b28 100644
--- a/vllm/model_executor/models/chameleon.py
+++ b/vllm/model_executor/models/chameleon.py
@@ -918,8 +918,6 @@ def forward(
 class ChameleonForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsPP, SupportsQuant
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
diff --git a/vllm/model_executor/models/clip.py b/vllm/model_executor/models/clip.py
index b8af3050990b..22f3ecad748e 100644
--- a/vllm/model_executor/models/clip.py
+++ b/vllm/model_executor/models/clip.py
@@ -784,7 +784,6 @@ class CLIPEmbeddingModel(nn.Module, SupportsMultiModal, SupportsQuant):
     is_pooling_model = True
 
     packed_modules_mapping = {"qkv_proj": ["q_proj", "k_proj", "v_proj"]}
-    merge_by_field_config = True
 
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
diff --git a/vllm/model_executor/models/cohere2_vision.py b/vllm/model_executor/models/cohere2_vision.py
index 139ccba9df6d..07dc7a01dc31 100644
--- a/vllm/model_executor/models/cohere2_vision.py
+++ b/vllm/model_executor/models/cohere2_vision.py
@@ -331,8 +331,6 @@ def get_replacement(item_idx: int):
     dummy_inputs=Cohere2VisionDummyInputsBuilder,
 )
 class Cohere2VisionForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             "model.vision_tower.": "vision_tower.",
diff --git a/vllm/model_executor/models/deepseek_ocr.py b/vllm/model_executor/models/deepseek_ocr.py
index a612ebd95628..1f07381c0cbd 100644
--- a/vllm/model_executor/models/deepseek_ocr.py
+++ b/vllm/model_executor/models/deepseek_ocr.py
@@ -344,8 +344,6 @@ def get_replacement_deepseek_vl2(item_idx: int):
     dummy_inputs=DeepseekOCRDummyInputsBuilder,
 )
 class DeepseekOCRForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             # map prefix for language backbone
diff --git a/vllm/model_executor/models/deepseek_vl2.py b/vllm/model_executor/models/deepseek_vl2.py
index 56c1a87a2540..9f8faf9ed91c 100644
--- a/vllm/model_executor/models/deepseek_vl2.py
+++ b/vllm/model_executor/models/deepseek_vl2.py
@@ -344,8 +344,6 @@ def _cached_apply_hf_processor(
     dummy_inputs=DeepseekVL2DummyInputsBuilder,
 )
 class DeepseekVLV2ForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             "language.": "language_model.",
diff --git a/vllm/model_executor/models/dots_ocr.py b/vllm/model_executor/models/dots_ocr.py
index 5cc2a48f26d6..da19d8fdb15e 100644
--- a/vllm/model_executor/models/dots_ocr.py
+++ b/vllm/model_executor/models/dots_ocr.py
@@ -690,8 +690,6 @@ def forward(
     dummy_inputs=DotsOCRDummyInputsBuilder,
 )
 class DotsOCRForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_substr={
             ".attn.qkv_proj.": ".attn.qkv.",
diff --git a/vllm/model_executor/models/ernie45_vl.py b/vllm/model_executor/models/ernie45_vl.py
index 81663dd7bbb4..3305b6a0e58f 100644
--- a/vllm/model_executor/models/ernie45_vl.py
+++ b/vllm/model_executor/models/ernie45_vl.py
@@ -1254,8 +1254,6 @@ def get_dummy_mm_data(
 class Ernie4_5_VLMoeForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsLoRA, SupportsPP, SupportsMRoPE
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
diff --git a/vllm/model_executor/models/fuyu.py b/vllm/model_executor/models/fuyu.py
index 269c36ab5b9c..8a7a3dd771c3 100644
--- a/vllm/model_executor/models/fuyu.py
+++ b/vllm/model_executor/models/fuyu.py
@@ -260,8 +260,6 @@ def get_replacement_fuyu(item_idx: int):
     dummy_inputs=FuyuDummyInputsBuilder,
 )
 class FuyuForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             "model.vision_embed_tokens.": "vision_embed_tokens.",
diff --git a/vllm/model_executor/models/gemma3_mm.py b/vllm/model_executor/models/gemma3_mm.py
index 43c69e5e1399..e8dec36a1c5b 100644
--- a/vllm/model_executor/models/gemma3_mm.py
+++ b/vllm/model_executor/models/gemma3_mm.py
@@ -483,8 +483,6 @@ def forward(self, vision_outputs: torch.Tensor):
 class Gemma3ForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
diff --git a/vllm/model_executor/models/gemma3n_mm.py b/vllm/model_executor/models/gemma3n_mm.py
index 6ae76976eb46..7036118ada08 100644
--- a/vllm/model_executor/models/gemma3n_mm.py
+++ b/vllm/model_executor/models/gemma3n_mm.py
@@ -463,7 +463,6 @@ def forward(
 class Gemma3nForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsTranscription
 ):
-    merge_by_field_config = True
     supported_languages = ISO639_1_SUPPORTED_LANGS
 
     packed_modules_mapping = {
diff --git a/vllm/model_executor/models/glm4_1v.py b/vllm/model_executor/models/glm4_1v.py
index 5ba3c0a35928..3cb53f2cbabe 100644
--- a/vllm/model_executor/models/glm4_1v.py
+++ b/vllm/model_executor/models/glm4_1v.py
@@ -1424,8 +1424,6 @@ def get_video_replacement_glm4v(item_idx: int):
 class Glm4vForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsLoRA, SupportsPP, SupportsMRoPE
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
diff --git a/vllm/model_executor/models/glm4v.py b/vllm/model_executor/models/glm4v.py
index 514082cf60ce..ec5af94e297c 100644
--- a/vllm/model_executor/models/glm4v.py
+++ b/vllm/model_executor/models/glm4v.py
@@ -561,8 +561,6 @@ def get_replacement(item_idx: int):
 class GLM4VForCausalLM(
     ChatGLMBaseModel, SupportsMultiModal, SupportsLoRA, SupportsPP, SupportsMRoPE
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "query_key_value": ["query_key_value"],
         "dense_h_to_4h": ["dense_h_to_4h"],
diff --git a/vllm/model_executor/models/granite_speech.py b/vllm/model_executor/models/granite_speech.py
index accf7e6ef2f4..a4e50f408628 100644
--- a/vllm/model_executor/models/granite_speech.py
+++ b/vllm/model_executor/models/granite_speech.py
@@ -564,7 +564,6 @@ class GraniteSpeechForConditionalGeneration(
     SupportsLoRA,
     SupportsTranscription,
 ):
-    merge_by_field_config = True
     supported_languages = ISO639_1_SUPPORTED_LANGS
 
     packed_modules_mapping = {
diff --git a/vllm/model_executor/models/hunyuan_vision.py b/vllm/model_executor/models/hunyuan_vision.py
index 5aef09ca9c25..52ce9564c8d7 100644
--- a/vllm/model_executor/models/hunyuan_vision.py
+++ b/vllm/model_executor/models/hunyuan_vision.py
@@ -786,7 +786,6 @@ class HunYuanVLForConditionalGeneration(
     SupportsQuant,
     SupportsXDRoPE,
 ):
-    merge_by_field_config = True
     multimodal_cpu_fields = {"image_grid_thw"}
 
     # To ensure correct weight loading and mapping.
diff --git a/vllm/model_executor/models/hyperclovax_vision.py b/vllm/model_executor/models/hyperclovax_vision.py
index db46353efde5..3a083870e4b5 100644
--- a/vllm/model_executor/models/hyperclovax_vision.py
+++ b/vllm/model_executor/models/hyperclovax_vision.py
@@ -592,8 +592,6 @@ def build_mlp(
     dummy_inputs=HCXVisionDummyInputsBuilder,
 )
 class HCXVisionForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
diff --git a/vllm/model_executor/models/idefics3.py b/vllm/model_executor/models/idefics3.py
index 7c3933c6feb7..0eed46448786 100644
--- a/vllm/model_executor/models/idefics3.py
+++ b/vllm/model_executor/models/idefics3.py
@@ -576,8 +576,6 @@ def forward(
     dummy_inputs=Idefics3DummyInputsBuilder,
 )
 class Idefics3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsLoRA):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
diff --git a/vllm/model_executor/models/interfaces.py b/vllm/model_executor/models/interfaces.py
index 01b3e7827424..416ab236cd18 100644
--- a/vllm/model_executor/models/interfaces.py
+++ b/vllm/model_executor/models/interfaces.py
@@ -78,9 +78,9 @@ class SupportsMultiModal(Protocol):
     `multimodal_config.mm_encoder_tp_mode="data"`.
     """
 
-    merge_by_field_config: ClassVar[bool] = True
+    merge_by_field_config: ClassVar[bool | None] = None
     """
-    A flag that indicates which implementation of
+    [DEPRECATED] A flag that indicates which implementation of
     `vllm.multimodal.utils.group_mm_kwargs_by_modality` to use.
     """
 
@@ -260,7 +260,26 @@ def supports_multimodal(model: object) -> TypeIs[SupportsMultiModal]: ...
 def supports_multimodal(
     model: type[object] | object,
 ) -> TypeIs[type[SupportsMultiModal]] | TypeIs[SupportsMultiModal]:
-    return getattr(model, "supports_multimodal", False)
+    res = getattr(model, "supports_multimodal", False)
+
+    if res:
+        # We can remove this starting from v0.14
+        merge_by_field_config = getattr(model, "merge_by_field_config", None)
+        if merge_by_field_config is False:
+            raise ValueError(
+                "`merge_by_field_config=False` is no longer effective, "
+                "please update your model to consider the new batching logic "
+                "in `group_mm_kwargs_by_modality` (refer to "
+                "https://github.com/vllm-project/vllm/issues/26149), "
+                "and then remove the override from your model."
+            )
+        if merge_by_field_config is True:
+            logger.warning_once(
+                "`merge_by_field_config=True` is redundant, "
+                "please remove the override from your model."
+            )
+
+    return res
 
 
 def supports_multimodal_raw_input_only(model: type[object] | object) -> bool:
diff --git a/vllm/model_executor/models/interns1.py b/vllm/model_executor/models/interns1.py
index c2195fd0cb88..18985cefbf5e 100644
--- a/vllm/model_executor/models/interns1.py
+++ b/vllm/model_executor/models/interns1.py
@@ -509,8 +509,6 @@ def get_replacement_interns1_video(item_idx: int):
 class InternS1ForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA
 ):
-    merge_by_field_config = True
-
     # To ensure correct weight loading and mapping.
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
diff --git a/vllm/model_executor/models/internvl.py b/vllm/model_executor/models/internvl.py
index fccddf3a6b29..15f7d4f418e4 100644
--- a/vllm/model_executor/models/internvl.py
+++ b/vllm/model_executor/models/internvl.py
@@ -1074,8 +1074,6 @@ def get_video_replacement_internvl(item_idx: int):
     dummy_inputs=InternVLDummyInputsBuilder,
 )
 class InternVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA):
-    merge_by_field_config = True
-
     supports_encoder_tp_data = True
 
     @classmethod
diff --git a/vllm/model_executor/models/keye.py b/vllm/model_executor/models/keye.py
index 09acf8372e16..f31da0ee302b 100644
--- a/vllm/model_executor/models/keye.py
+++ b/vllm/model_executor/models/keye.py
@@ -1292,8 +1292,6 @@ def _get_mm_fields_config(
 
 
 class BaseKeyeModule(nn.Module):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
diff --git a/vllm/model_executor/models/kimi_vl.py b/vllm/model_executor/models/kimi_vl.py
index 8167b82f3233..85267ccda8a9 100644
--- a/vllm/model_executor/models/kimi_vl.py
+++ b/vllm/model_executor/models/kimi_vl.py
@@ -298,8 +298,6 @@ def get_replacement(item_idx: int):
     dummy_inputs=KimiVLDummyInputsBuilder,
 )
 class KimiVLForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     supports_encoder_tp_data = True
 
     @classmethod
diff --git a/vllm/model_executor/models/llava.py b/vllm/model_executor/models/llava.py
index c1fb2d4f4af7..66a327bb7603 100644
--- a/vllm/model_executor/models/llava.py
+++ b/vllm/model_executor/models/llava.py
@@ -506,8 +506,6 @@ def init_vision_tower_for_llava(
     dummy_inputs=LlavaDummyInputsBuilder,
 )
 class LlavaForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
diff --git a/vllm/model_executor/models/llava_next.py b/vllm/model_executor/models/llava_next.py
index b995cac47ac1..526846d0d981 100644
--- a/vllm/model_executor/models/llava_next.py
+++ b/vllm/model_executor/models/llava_next.py
@@ -223,8 +223,6 @@ def _get_mm_fields_config(
     dummy_inputs=LlavaDummyInputsBuilder,
 )
 class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             # mapping for new names in checkpoint saved after transformers v4.52
diff --git a/vllm/model_executor/models/llava_next_video.py b/vllm/model_executor/models/llava_next_video.py
index 902c598c226f..cd55cfec6cde 100644
--- a/vllm/model_executor/models/llava_next_video.py
+++ b/vllm/model_executor/models/llava_next_video.py
@@ -299,8 +299,6 @@ def forward(self, image_features: torch.Tensor) -> torch.Tensor:
     dummy_inputs=LlavaNextVideoDummyInputsBuilder,
 )
 class LlavaNextVideoForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             # mapping for new names in checkpoint saved after transformers v4.52
diff --git a/vllm/model_executor/models/llava_onevision.py b/vllm/model_executor/models/llava_onevision.py
index 4e243ade6835..5aa8de7dc252 100644
--- a/vllm/model_executor/models/llava_onevision.py
+++ b/vllm/model_executor/models/llava_onevision.py
@@ -479,8 +479,6 @@ def forward(self, image_features: torch.Tensor) -> torch.Tensor:
     dummy_inputs=LlavaOnevisionDummyInputsBuilder,
 )
 class LlavaOnevisionForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             # mapping for new names in checkpoint saved after transformers v4.52
diff --git a/vllm/model_executor/models/midashenglm.py b/vllm/model_executor/models/midashenglm.py
index d9b23811730d..2d506978d266 100644
--- a/vllm/model_executor/models/midashenglm.py
+++ b/vllm/model_executor/models/midashenglm.py
@@ -683,8 +683,6 @@ def get_replacement_midashenglm(item_idx: int):
     dummy_inputs=MiDashengLMDummyInputsBuilder,
 )
 class MiDashengLMModel(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
diff --git a/vllm/model_executor/models/minicpmv.py b/vllm/model_executor/models/minicpmv.py
index 6d0ebf5c9825..c45bdf95e748 100644
--- a/vllm/model_executor/models/minicpmv.py
+++ b/vllm/model_executor/models/minicpmv.py
@@ -1003,8 +1003,6 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):
     instantiated.
     """
 
-    merge_by_field_config = True
-
     supports_encoder_tp_data = True
 
     @classmethod
diff --git a/vllm/model_executor/models/minimax_vl_01.py b/vllm/model_executor/models/minimax_vl_01.py
index 0939a72ba53e..e480454953df 100644
--- a/vllm/model_executor/models/minimax_vl_01.py
+++ b/vllm/model_executor/models/minimax_vl_01.py
@@ -179,8 +179,6 @@ def _get_mm_fields_config(
     dummy_inputs=MiniMaxVL01DummyInputsBuilder,
 )
 class MiniMaxVL01ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
diff --git a/vllm/model_executor/models/mistral3.py b/vllm/model_executor/models/mistral3.py
index 1ddb470a0f93..e9161e69e731 100644
--- a/vllm/model_executor/models/mistral3.py
+++ b/vllm/model_executor/models/mistral3.py
@@ -423,8 +423,6 @@ def init_vision_tower_for_llava(
 class Mistral3ForConditionalGeneration(
     nn.Module, SupportsLoRA, SupportsMultiModal, SupportsPP
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
diff --git a/vllm/model_executor/models/mllama4.py b/vllm/model_executor/models/mllama4.py
index 286859d188d3..e944c0ee38aa 100644
--- a/vllm/model_executor/models/mllama4.py
+++ b/vllm/model_executor/models/mllama4.py
@@ -741,8 +741,6 @@ class Llama4ForConditionalGeneration(
     SupportsEagle3,
     SupportsLoRA,
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
diff --git a/vllm/model_executor/models/molmo.py b/vllm/model_executor/models/molmo.py
index 7b53299cccbe..a6cd9ad16c18 100644
--- a/vllm/model_executor/models/molmo.py
+++ b/vllm/model_executor/models/molmo.py
@@ -1354,8 +1354,6 @@ def get_insertion_molmo(item_idx: int):
 class MolmoForCausalLM(
     nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA, SupportsQuant
 ):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_substr={
             # vision backbone mapping
diff --git a/vllm/model_executor/models/nano_nemotron_vl.py b/vllm/model_executor/models/nano_nemotron_vl.py
index c4198d36b392..6dfab595e5b9 100644
--- a/vllm/model_executor/models/nano_nemotron_vl.py
+++ b/vllm/model_executor/models/nano_nemotron_vl.py
@@ -1116,8 +1116,6 @@ def get_dummy_mm_data(
 class NemotronH_Nano_VL_V2(
     nn.Module, HasInnerState, IsHybrid, SupportsMultiModal, SupportsMultiModalPruning
 ):
-    merge_by_field_config = True
-
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):
diff --git a/vllm/model_executor/models/nemotron_vl.py b/vllm/model_executor/models/nemotron_vl.py
index a57668b21fb8..391980fc61f9 100644
--- a/vllm/model_executor/models/nemotron_vl.py
+++ b/vllm/model_executor/models/nemotron_vl.py
@@ -358,8 +358,6 @@ def get_image_processor(self, **kwargs: object):
     dummy_inputs=BaseInternVLDummyInputsBuilder[NemotronVLProcessingInfo],
 )
 class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA):
-    merge_by_field_config = True
-
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):
diff --git a/vllm/model_executor/models/opencua.py b/vllm/model_executor/models/opencua.py
index b92f0c9dac32..76a2d1cc242c 100644
--- a/vllm/model_executor/models/opencua.py
+++ b/vllm/model_executor/models/opencua.py
@@ -201,7 +201,6 @@ def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
     dummy_inputs=OpenCUADummyInputsBuilder,
 )
 class OpenCUAForConditionalGeneration(Qwen2_5_VLForConditionalGeneration):
-    merge_by_field_config = True
     multimodal_cpu_fields = {"image_grid_thw"}
 
     packed_modules_mapping = {
diff --git a/vllm/model_executor/models/ovis.py b/vllm/model_executor/models/ovis.py
index a0fab820720f..0691bbc615be 100644
--- a/vllm/model_executor/models/ovis.py
+++ b/vllm/model_executor/models/ovis.py
@@ -414,8 +414,6 @@ def get_replacement_ovis(item_idx: int):
     dummy_inputs=OvisDummyInputsBuilder,
 )
 class Ovis(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):
diff --git a/vllm/model_executor/models/ovis2_5.py b/vllm/model_executor/models/ovis2_5.py
index 85f37cfea10b..0ad22aab748e 100644
--- a/vllm/model_executor/models/ovis2_5.py
+++ b/vllm/model_executor/models/ovis2_5.py
@@ -456,8 +456,6 @@ def get_replacement_ovis(item_idx, modality: str):
     dummy_inputs=Ovis2_5DummyInputsBuilder,
 )
 class Ovis2_5(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
         config = vllm_config.model_config.hf_config
diff --git a/vllm/model_executor/models/paddleocr_vl.py b/vllm/model_executor/models/paddleocr_vl.py
index 1df5ff62fa5b..9703a5b417d0 100644
--- a/vllm/model_executor/models/paddleocr_vl.py
+++ b/vllm/model_executor/models/paddleocr_vl.py
@@ -1103,8 +1103,6 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
     dummy_inputs=PaddleOCRVLDummyInputsBuilder,
 )
 class PaddleOCRVLForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsMRoPE):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             "model.": "language_model.model.",
diff --git a/vllm/model_executor/models/paligemma.py b/vllm/model_executor/models/paligemma.py
index 9fa32f01d37a..67240c6e7124 100644
--- a/vllm/model_executor/models/paligemma.py
+++ b/vllm/model_executor/models/paligemma.py
@@ -251,8 +251,6 @@ def apply(
     dummy_inputs=PaliGemmaDummyInputsBuilder,
 )
 class PaliGemmaForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
diff --git a/vllm/model_executor/models/phi3v.py b/vllm/model_executor/models/phi3v.py
index 384572217bc1..b7ae548069f2 100644
--- a/vllm/model_executor/models/phi3v.py
+++ b/vllm/model_executor/models/phi3v.py
@@ -562,8 +562,6 @@ def _apply_prompt_updates(
     dummy_inputs=Phi3VDummyInputsBuilder,
 )
 class Phi3VForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, SupportsQuant):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             "model.vision_embed_tokens.wte": "embed_tokens",
diff --git a/vllm/model_executor/models/phi4mm.py b/vllm/model_executor/models/phi4mm.py
index 8425549a7bd2..179d5df869be 100644
--- a/vllm/model_executor/models/phi4mm.py
+++ b/vllm/model_executor/models/phi4mm.py
@@ -984,8 +984,6 @@ class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):
     Implements the Phi-4-multimodal-instruct model in vLLM.
     """
 
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": [
             "qkv_proj",
diff --git a/vllm/model_executor/models/pixtral.py b/vllm/model_executor/models/pixtral.py
index cad241842cd3..faf2d80d24bb 100644
--- a/vllm/model_executor/models/pixtral.py
+++ b/vllm/model_executor/models/pixtral.py
@@ -365,8 +365,6 @@ def _cached_apply_hf_processor(
     dummy_inputs=PixtralDummyInputsBuilder,
 )
 class PixtralForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):
diff --git a/vllm/model_executor/models/qwen2_5_omni_thinker.py b/vllm/model_executor/models/qwen2_5_omni_thinker.py
index 1ce0fb4e4d93..3438406c4fac 100644
--- a/vllm/model_executor/models/qwen2_5_omni_thinker.py
+++ b/vllm/model_executor/models/qwen2_5_omni_thinker.py
@@ -773,8 +773,6 @@ class Qwen2_5OmniThinkerForConditionalGeneration(
     SupportsMRoPE,
     Qwen2_5OmniConditionalGenerationMixin,
 ):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             "thinker.lm_head.": "language_model.lm_head.",
diff --git a/vllm/model_executor/models/qwen2_5_vl.py b/vllm/model_executor/models/qwen2_5_vl.py
index cb521ebdf0af..488af192bfad 100644
--- a/vllm/model_executor/models/qwen2_5_vl.py
+++ b/vllm/model_executor/models/qwen2_5_vl.py
@@ -1039,7 +1039,6 @@ class Qwen2_5_VLForConditionalGeneration(
     SupportsMultiModalPruning,
     SupportsMRoPE,
 ):
-    merge_by_field_config = True
     multimodal_cpu_fields = {"image_grid_thw", "video_grid_thw"}
 
     packed_modules_mapping = {
diff --git a/vllm/model_executor/models/qwen2_audio.py b/vllm/model_executor/models/qwen2_audio.py
index 7e883a393aa8..f84ddfa84f6a 100644
--- a/vllm/model_executor/models/qwen2_audio.py
+++ b/vllm/model_executor/models/qwen2_audio.py
@@ -313,8 +313,6 @@ def get_replacement_qwen2_audio(item_idx: int):
     dummy_inputs=Qwen2AudioDummyInputsBuilder,
 )
 class Qwen2AudioForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("audio"):
diff --git a/vllm/model_executor/models/qwen2_vl.py b/vllm/model_executor/models/qwen2_vl.py
index b74876849841..9da5080f8430 100644
--- a/vllm/model_executor/models/qwen2_vl.py
+++ b/vllm/model_executor/models/qwen2_vl.py
@@ -1131,7 +1131,6 @@ def _get_mm_fields_config(
 class Qwen2VLForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsLoRA, SupportsPP, SupportsMRoPE
 ):
-    merge_by_field_config = True
     multimodal_cpu_fields = {"image_grid_thw", "video_grid_thw"}
 
     # To ensure correct weight loading and mapping.
diff --git a/vllm/model_executor/models/qwen3_omni_moe_thinker.py b/vllm/model_executor/models/qwen3_omni_moe_thinker.py
index e6979211b707..dbe7bcd07576 100755
--- a/vllm/model_executor/models/qwen3_omni_moe_thinker.py
+++ b/vllm/model_executor/models/qwen3_omni_moe_thinker.py
@@ -1131,8 +1131,6 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(
     SupportsMRoPE,
     Qwen3OmniMoeConditionalGenerationMixin,
 ):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             "thinker.lm_head.": "language_model.lm_head.",
diff --git a/vllm/model_executor/models/qwen3_vl.py b/vllm/model_executor/models/qwen3_vl.py
index 58721303dfc8..a5b10c95872d 100644
--- a/vllm/model_executor/models/qwen3_vl.py
+++ b/vllm/model_executor/models/qwen3_vl.py
@@ -1190,7 +1190,6 @@ class Qwen3VLForConditionalGeneration(
     SupportsMRoPE,
     SupportsEagle3,
 ):
-    merge_by_field_config = True
     multimodal_cpu_fields = {"image_grid_thw", "video_grid_thw"}
 
     packed_modules_mapping = {
diff --git a/vllm/model_executor/models/qwen_vl.py b/vllm/model_executor/models/qwen_vl.py
index 55680b8e7ddf..caac14716782 100644
--- a/vllm/model_executor/models/qwen_vl.py
+++ b/vllm/model_executor/models/qwen_vl.py
@@ -703,8 +703,6 @@ def _get_prompt_updates(
 class QwenVLForConditionalGeneration(
     QWenBaseModel, SupportsPP, SupportsLoRA, SupportsMultiModal
 ):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "c_attn": ["c_attn"],
         "gate_up_proj": [
diff --git a/vllm/model_executor/models/siglip.py b/vllm/model_executor/models/siglip.py
index 9db1423d98e0..2600dc1c9f79 100644
--- a/vllm/model_executor/models/siglip.py
+++ b/vllm/model_executor/models/siglip.py
@@ -989,7 +989,6 @@ class SiglipEmbeddingModel(nn.Module, SupportsMultiModal, SupportsQuant):
     is_pooling_model = True
 
     packed_modules_mapping = {"qkv_proj": ["q_proj", "k_proj", "v_proj"]}
-    merge_by_field_config = True
 
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
diff --git a/vllm/model_executor/models/skyworkr1v.py b/vllm/model_executor/models/skyworkr1v.py
index 55c25ce6190f..f95fbffc1d0b 100644
--- a/vllm/model_executor/models/skyworkr1v.py
+++ b/vllm/model_executor/models/skyworkr1v.py
@@ -647,8 +647,6 @@ def get_replacement_skyworkr1v(item_idx: int):
     dummy_inputs=SkyworkR1VDummyInputsBuilder,
 )
 class SkyworkR1VChatModel(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     @classmethod
     def get_placeholder_str(cls, modality: str, i: int) -> str | None:
         if modality.startswith("image"):
diff --git a/vllm/model_executor/models/step3_vl.py b/vllm/model_executor/models/step3_vl.py
index 3e55ada0ed2e..e5038e56a270 100644
--- a/vllm/model_executor/models/step3_vl.py
+++ b/vllm/model_executor/models/step3_vl.py
@@ -916,8 +916,6 @@ def forward(
     dummy_inputs=Step3VLDummyInputsBuilder,
 )
 class Step3VLForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={
             "model.": "language_model.model.",
diff --git a/vllm/model_executor/models/tarsier.py b/vllm/model_executor/models/tarsier.py
index 4d310712f303..7e82a4d725a6 100644
--- a/vllm/model_executor/models/tarsier.py
+++ b/vllm/model_executor/models/tarsier.py
@@ -400,8 +400,6 @@ def _get_layer_index(feature_layer_index: int, num_hidden_layers_total: int) ->
     dummy_inputs=TarsierDummyInputsBuilder,
 )
 class TarsierForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
diff --git a/vllm/model_executor/models/terratorch.py b/vllm/model_executor/models/terratorch.py
index 9f34090e3107..402081a70631 100644
--- a/vllm/model_executor/models/terratorch.py
+++ b/vllm/model_executor/models/terratorch.py
@@ -227,7 +227,6 @@ def apply(
     dummy_inputs=TerratorchInputBuilder,
 )
 class Terratorch(nn.Module, IsAttentionFree, SupportsMultiModal):
-    merge_by_field_config = True
     supports_multimodal_raw_input_only = True
     is_pooling_model = True
 
diff --git a/vllm/model_executor/models/transformers/multimodal.py b/vllm/model_executor/models/transformers/multimodal.py
index ccf605371987..9d77dee2810c 100644
--- a/vllm/model_executor/models/transformers/multimodal.py
+++ b/vllm/model_executor/models/transformers/multimodal.py
@@ -264,7 +264,7 @@ def apply(
 
 class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
     supports_multimodal_raw_input_only = True
-    merge_by_field_config = True
+
     # Backwards compatibility for prev released models. State dicts back then
     # had different formats and cannot be loaded with `AutoModel` mapping as is
     hf_to_vllm_mapper = WeightsMapper(
diff --git a/vllm/model_executor/models/ultravox.py b/vllm/model_executor/models/ultravox.py
index 2444159b2ad6..32a2ba1ef38f 100644
--- a/vllm/model_executor/models/ultravox.py
+++ b/vllm/model_executor/models/ultravox.py
@@ -498,8 +498,6 @@ def forward(
     dummy_inputs=UltravoxDummyInputsBuilder,
 )
 class UltravoxModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA):
-    merge_by_field_config = True
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
diff --git a/vllm/model_executor/models/voxtral.py b/vllm/model_executor/models/voxtral.py
index 45f8fa079c71..7b408248ec74 100644
--- a/vllm/model_executor/models/voxtral.py
+++ b/vllm/model_executor/models/voxtral.py
@@ -330,8 +330,6 @@ def _get_data_parser(self) -> MultiModalDataParser:
 class VoxtralForConditionalGeneration(
     nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA, SupportsTranscription
 ):
-    merge_by_field_config = True
-
     supported_languages = ISO639_1_SUPPORTED_LANGS
 
     packed_modules_mapping = {
diff --git a/vllm/model_executor/models/whisper.py b/vllm/model_executor/models/whisper.py
index 0daf6bda61cc..b2feff133515 100644
--- a/vllm/model_executor/models/whisper.py
+++ b/vllm/model_executor/models/whisper.py
@@ -775,7 +775,6 @@ def _get_prompt_updates(
 class WhisperForConditionalGeneration(
     nn.Module, SupportsTranscription, SupportsMultiModal
 ):
-    merge_by_field_config = True
     packed_modules_mapping = {
         "self_attn.qkv_proj": [
             "self_attn.q_proj",
diff --git a/vllm/multimodal/utils.py b/vllm/multimodal/utils.py
index f8e8847e8e60..9c5e3fb2b32a 100644
--- a/vllm/multimodal/utils.py
+++ b/vllm/multimodal/utils.py
@@ -426,7 +426,6 @@ def group_mm_kwargs_by_modality(
     Yields:
         A tuple `(modality, num_items, grouped_kwargs)`.
     """
-    # TODO: After v0.13, remove merge_by_field_config attribute from model impls
     if merge_by_field_config is not None:
         logger.warning_once(
             "The `merge_by_field_config` argument of `group_mm_kwargs_by_modality` "