diff --git a/swift/template/templates/mimo.py b/swift/template/templates/mimo.py
index c4fd1f54b4..e8e2c21016 100644
--- a/swift/template/templates/mimo.py
+++ b/swift/template/templates/mimo.py
@@ -1,7 +1,6 @@
 # Copyright (c) ModelScope Contributors. All rights reserved.
 import torch
 from dataclasses import dataclass, field
-from qwen_vl_utils import fetch_image, fetch_video
 from typing import Any, Dict, List, Optional
 
 from ..base import Template
@@ -32,6 +31,8 @@ class MiMoV2Template(Template):
     norm_bbox = 'none'
 
     def replace_tag(self, media_type, index, inputs: StdTemplateInputs) -> List[Context]:
+        from qwen_vl_utils import fetch_image, fetch_video
+
         assert media_type in {'image', 'video'}
         kwargs = {'image_patch_size': self.processor.image_processor.patch_size}
         if media_type == 'image':
diff --git a/tests/general/test_optional_template_dependencies.py b/tests/general/test_optional_template_dependencies.py
new file mode 100644
index 0000000000..9857224df0
--- /dev/null
+++ b/tests/general/test_optional_template_dependencies.py
@@ -0,0 +1,29 @@
+# Copyright (c) ModelScope Contributors. All rights reserved.
+import subprocess
+import sys
+import unittest
+
+
+class TestOptionalTemplateDependencies(unittest.TestCase):
+
+    def test_template_import_without_qwen_vl_utils(self):
+        code = """
+import builtins
+
+original_import = builtins.__import__
+
+
+def import_without_qwen_vl_utils(name, *args, **kwargs):
+    if name == 'qwen_vl_utils' or name.startswith('qwen_vl_utils.'):
+        raise ModuleNotFoundError("No module named 'qwen_vl_utils'")
+    return original_import(name, *args, **kwargs)
+
+
+builtins.__import__ = import_without_qwen_vl_utils
+import swift.template
+"""
+        subprocess.run([sys.executable, '-c', code], check=True)
+
+
+if __name__ == '__main__':
+    unittest.main()