diff --git a/auto_round/compressors/mllm/compressor.py b/auto_round/compressors/mllm/compressor.py
index 2cfa457b7..94210388f 100644
--- a/auto_round/compressors/mllm/compressor.py
+++ b/auto_round/compressors/mllm/compressor.py
@@ -27,6 +27,7 @@
 from auto_round.logger import logger
 from auto_round.schemes import QuantizationScheme
 from auto_round.special_model_handler import (
+    MISTRAL_3_2_MODELS,
     NOT_SUPPORT_ONLY_TEXT_MODELS,
     SUPPORT_ONLY_TEXT_MODELS,
     _handle_special_model,
@@ -164,6 +165,7 @@ def __init__(
         seed: int = 42,
         **kwargs,
     ):
+
         extra_data_dir = kwargs.pop("extra_data_dir", None)
         template = kwargs.pop("template", None)
 
@@ -189,7 +191,7 @@ def __init__(
 
         if model.config.model_type == "llava" and isinstance(model, PreTrainedModel):
             template = "default"
-        if hasattr(model, "name_or_path") and "Mistral-Small-3.2" in model.name_or_path:
+        if hasattr(model, "name_or_path") and any([name in model.name_or_path for name in MISTRAL_3_2_MODELS]):
             template = "mistral3_2"
         if iters > 0:
             self.template = template if template is not None else model.config.model_type
diff --git a/auto_round/special_model_handler.py b/auto_round/special_model_handler.py
index d0fa8c962..a999b358b 100644
--- a/auto_round/special_model_handler.py
+++ b/auto_round/special_model_handler.py
@@ -39,6 +39,8 @@
 
 CONVERT_EXPERT_TO_LINEAR_MODELS = ["llama4", "gpt_oss"]
 
+MISTRAL_3_2_MODELS = ["Mistral-Small-3.2", "Magistral-Small", "Devstral-Small"]
+
 
 def _get_moe_converter(config):
     # Dispatch table for model_type to replacement_info functions
diff --git a/auto_round/utils/model.py b/auto_round/utils/model.py
index 427ec7f9c..1aab07983 100644
--- a/auto_round/utils/model.py
+++ b/auto_round/utils/model.py
@@ -307,6 +307,8 @@ def mllm_load_model(
     model_dtype: str = None,
     **kwargs,
 ):
+    from auto_round.special_model_handler import MISTRAL_3_2_MODELS
+
     assert platform.lower() in [
         "hf",
         "model_scope",
@@ -410,7 +412,7 @@ def mllm_load_model(
                 else:
                     raise
 
-            if "Mistral-Small-3.2" in pretrained_model_name_or_path:
+            if any([name in model.name_or_path for name in MISTRAL_3_2_MODELS]):
                 from mistral_common.tokens.tokenizers.mistral import MistralTokenizer  # pylint: disable=E0401
 
                 if os.path.isdir(pretrained_model_name_or_path):