[model] support XiaomiMiMo/MiMo-V2.5

swift/model/constant.py

@@ -154,6 +154,7 @@ class MLLMModelType:
     ovis2 = 'ovis2'
     ovis2_5 = 'ovis2_5'
     midashenglm = 'midashenglm'
+    mimo_v2 = 'mimo_v2'
 
     chatglm4v = 'chatglm4v'
     glm4v = 'glm4v'

swift/model/model_arch.py

@@ -87,6 +87,7 @@ class MLLMModelArch:
     keye_vl = 'keye_vl'
 
     midashenglm = 'midashenglm'
+    mimo_v2 = 'mimo_v2'
     step_audio2_mini = 'step_audio2_mini'
     hunyuan_vl = 'hunyuan_vl'
     step3_vl = 'step3_vl'
@@ -787,6 +788,14 @@ def register_model_arch(model_arch: ModelKeys, *, exist_ok: bool = False) -> Non
         vision_tower='model.visual',
     ))
 
+register_model_arch(
+    MultiModelKeys(
+        MLLMModelArch.mimo_v2,
+        language_model=['model', 'lm_head'],
+        aligner='visual.merger',
+        vision_tower=['visual', 'audio_encoder', 'speech_embeddings'],
+    ))
+
 
 def get_model_arch(arch_name: Optional[str]) -> Optional[MultiModelKeys]:
     return MODEL_ARCH_MAPPING.get(arch_name)

swift/model/models/qwen.py

@@ -867,6 +867,28 @@ def get_model(self, model_dir: str, *args, **kwargs) -> PreTrainedModel:
         tags=['vision', 'video']))
 
 
+class MiMoV2Loader(Qwen2VLLoader):
+
+    def get_model(self, model_dir: str, config, processor, model_kwargs) -> PreTrainedModel:
+        model = ModelLoader.get_model(self, model_dir, config, processor, model_kwargs)
+        patch_get_input_embeddings(model.visual, 'patch_embed')
+        return model
+
+
+register_model(
+    ModelMeta(
+        MLLMModelType.mimo_v2, [
+            ModelGroup([
+                Model('XiaomiMiMo/MiMo-V2.5', 'XiaomiMiMo/MiMo-V2.5'),
+            ], TemplateType.mimo_v2),
+        ],
+        MiMoV2Loader,
+        model_arch=ModelArch.mimo_v2,
+        architectures=['MiMoV2ForCausalLM'],
+        requires=['transformers>=4.49', 'qwen_vl_utils>=0.0.6', 'decord'],
+        tags=['vision', 'video']))
+
+
 def patch_Qwen3VLMoeTextExperts_dtype():
     from transformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe import Qwen3VLMoeTextExperts
     if hasattr(Qwen3VLMoeTextExperts, '_patch'):

swift/template/constant.py

@@ -150,6 +150,7 @@ class MLLMTemplateType:
     ovis2 = 'ovis2'
     ovis2_5 = 'ovis2_5'
     mimo_vl = 'mimo_vl'
+    mimo_v2 = 'mimo_v2'
     midashenglm = 'midashenglm'
 
     llama3_1_omni = 'llama3_1_omni'

swift/template/templates/qwen.py

@@ -505,6 +505,50 @@ class Qwen2_5VLTemplate(Qwen2VLTemplate):
         default_system='You are MiMo, an AI assistant developed by Xiaomi.'))
 
 
+class MiMoV2Template(Qwen2_5VLTemplate):
+    """Template for XiaomiMiMo/MiMo-V2.5.
+
+    Differences from Qwen2_5VLTemplate:
+    - MiMo-V2.5 does not use 3D rope position IDs (no get_rope_index).
+    - Video key is named 'video_pixel_values' instead of 'pixel_values_videos'.
+    - Supports thinking mode with <think>...</think> tags.
+    """
+
+    def _get_position_ids(self, inputs: Dict[str, Any]):
+        # MiMo-V2.5 uses standard rotary position embeddings,
+        # not 3D rope like Qwen2VL. No special position IDs needed.
+        return {}
+
+    def forward_context(self, model, inputs):
+        # Skip Qwen2VL-specific flash attention patching
+        return Template.forward_context(self, model, inputs)
+
+    def _post_encode(self, model, inputs: Dict[str, Any]) -> Dict[str, Any]:
+        if not self.is_training:
+            # During inference, rename key to match MiMo-V2.5 forward signature
+            if 'pixel_values_videos' in inputs:
+                inputs['video_pixel_values'] = inputs.pop('pixel_values_videos')
+            return inputs
+        # For training, compute embeddings manually
+        input_ids = inputs['input_ids']
+        base_model = self.get_base_model(model)
+        inputs_embeds = base_model.model.embed_tokens(input_ids)
+        inputs_embeds = self._get_inputs_embeds_hf(inputs_embeds, inputs, model.visual, self.processor, model.config)
+        return {'inputs_embeds': inputs_embeds}
+
+
+register_template(
+    QwenTemplateMeta(
+        MLLMTemplateType.mimo_v2,
+        template_cls=MiMoV2Template,
+        default_system='You are MiMo, a helpful AI assistant engineered by Xiaomi.',
+        is_thinking=True,
+        thinking_prefix='<think>\n',
+        non_thinking_prefix='<think>\n</think>\n\n',
+        history_thinking_prefix='<think>\n</think>\n\n',
+    ))
+
+
 class Qwen3VLTemplate(Qwen2VLTemplate):
     version = 'v3'