diff --git a/litellm/llms/deepseek/chat/transformation.py b/litellm/llms/deepseek/chat/transformation.py
index 80775889118..4c3eb55cb54 100644
--- a/litellm/llms/deepseek/chat/transformation.py
+++ b/litellm/llms/deepseek/chat/transformation.py
@@ -70,8 +70,10 @@ class DeepSeekChatConfig(OpenAIGPTConfig):
         # can run away, exhausting max_tokens during reasoning and returning empty content
         # (finish_reason="length", content:null). Preserve the value into the request body
         # when thinking is on so callers can bound effort (e.g. reasoning_effort="low").
-        # See https://api-docs.deepseek.com/guides/thinking_mode
-        if thinking_enabled and reasoning_effort is not None:
+        # Exclude the "none" sentinel: it is an OpenAI-style thinking-OFF switch, not a real
+        # effort value, so passing it alongside explicit thinking:enabled would be
+        # contradictory. See https://api-docs.deepseek.com/guides/thinking_mode
+        if thinking_enabled and reasoning_effort not in (None, "none"):
             optional_params["reasoning_effort"] = reasoning_effort
 
         return optional_params
diff --git a/tests/llm_translation/test_deepseek_reasoning_effort.py b/tests/llm_translation/test_deepseek_reasoning_effort.py
index f9d3f29d2ab..6e1bb8f8281 100644
--- a/tests/llm_translation/test_deepseek_reasoning_effort.py
+++ b/tests/llm_translation/test_deepseek_reasoning_effort.py
@@ -39,3 +39,11 @@ def test_explicit_thinking_disabled_wins_and_drops_effort():
 def test_no_reasoning_effort_leaves_no_key():
     out = _map({})
     assert out.get("reasoning_effort") is None
+
+
+def test_explicit_enabled_thinking_with_none_effort_not_contradictory():
+    # "none" is a thinking-OFF sentinel, not a real effort value: it must not be
+    # passed alongside explicit thinking:enabled (would contradict the provider).
+    out = _map({"thinking": {"type": "enabled"}, "reasoning_effort": "none"})
+    assert out["thinking"] == {"type": "enabled"}
+    assert "reasoning_effort" not in out