diff --git a/src/langbot/pkg/provider/modelmgr/requesters/litellmchat.py b/src/langbot/pkg/provider/modelmgr/requesters/litellmchat.py
index ae1904b25..5fc7b8834 100644
--- a/src/langbot/pkg/provider/modelmgr/requesters/litellmchat.py
+++ b/src/langbot/pkg/provider/modelmgr/requesters/litellmchat.py
@@ -164,6 +164,12 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
 
     _EMBEDDING_MODEL_HINTS = ('embedding', 'embed', 'bge-', 'e5-', 'm3e', 'gte-', 'text-embedding')
     _RERANK_MODEL_HINTS = ('rerank', 're-rank', 're_rank')
+    _QWEN_DEDICATED_THINKING_MODELS = frozenset(
+        {
+            'qwen3.7-max-preview',
+            'qwen3.7-max-2026-05-17',
+        }
+    )
     _INFERRED_EFFORT_PROVIDERS = frozenset(
         {
             'anthropic',
@@ -392,6 +398,14 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
                 return inferred_family
             return 'volcengine' if requester_name == 'volcark-chat-completions' else ''
 
+        # Bailian's compatible endpoint also hosts Kimi models. Keep those
+        # models on Kimi's ``thinking`` protocol instead of Qwen's
+        # ``enable_thinking`` protocol.
+        if requester_name == 'bailian-chat-completions':
+            inferred_family = self._infer_reasoning_family_from_model_name(model_name)
+            if inferred_family == 'kimi':
+                return inferred_family
+
         requester_family = self._REQUESTER_REASONING_FAMILIES.get(requester_name)
         if requester_family:
             return requester_family
@@ -428,7 +442,11 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
     @staticmethod
     def _is_dedicated_qwen_thinking_model(model_name: str) -> bool:
         normalized_name = model_name.lower().rsplit('/', 1)[-1]
-        return normalized_name.startswith('qwq') or '-thinking' in normalized_name
+        return (
+            normalized_name in LiteLLMRequester._QWEN_DEDICATED_THINKING_MODELS
+            or normalized_name.startswith('qwq')
+            or '-thinking' in normalized_name
+        )
 
     def _known_reasoning_levels(self, model_name: str, family: str) -> list[str] | None:
         normalized_name = model_name.lower().rsplit('/', 1)[-1]
diff --git a/tests/unit_tests/provider/test_reasoning_control.py b/tests/unit_tests/provider/test_reasoning_control.py
index dabc642ad..2281037b1 100644
--- a/tests/unit_tests/provider/test_reasoning_control.py
+++ b/tests/unit_tests/provider/test_reasoning_control.py
@@ -237,6 +237,43 @@ def test_qwen_mixed_and_dedicated_thinking_profiles(monkeypatch):
     assert dedicated['levels'] == ['provider_default']
 
 
+@pytest.mark.parametrize('model_name', ['qwen3.7-max-preview', 'qwen3.7-max-2026-05-17'])
+def test_qwen_dedicated_thinking_release_models_are_not_toggleable(model_name, monkeypatch):
+    request = _requester('openai', 'bailian-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name=model_name, abilities=[]))
+
+    assert capabilities['levels'] == ['provider_default']
+
+
+@pytest.mark.parametrize(
+    ('model_name', 'expected_levels'),
+    [
+        ('kimi-k2.6', ['provider_default', 'disabled', 'enabled']),
+        ('kimi-k2.5', ['provider_default', 'disabled', 'enabled']),
+        ('kimi-k2.7-code', ['provider_default']),
+        ('kimi-k2-thinking', ['provider_default']),
+    ],
+)
+def test_bailian_kimi_profiles_use_kimi_model_rules(model_name, expected_levels, monkeypatch):
+    request = _requester('openai', 'bailian-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name=model_name, abilities=[]))
+
+    assert capabilities['levels'] == expected_levels
+
+
+def test_bailian_kimi_uses_thinking_protocol_instead_of_qwen_protocol(monkeypatch):
+    request = _requester('openai', 'bailian-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    assert request._build_reasoning_args(_runtime_model(request, 'disabled', name='kimi-k2.6')) == {
+        'extra_body': {'thinking': {'type': 'disabled'}}
+    }
+
+
 def test_doubao_exposes_documented_effort_range(monkeypatch):
     request = _requester('openai', 'doubao-chat-completions')
     monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)