diff --git a/litellm/llms/bedrock/chat/converse_transformation.py b/litellm/llms/bedrock/chat/converse_transformation.py
index c38b3593465..08f44d5b1b1 100644
--- a/litellm/llms/bedrock/chat/converse_transformation.py
+++ b/litellm/llms/bedrock/chat/converse_transformation.py
@@ -493,6 +493,23 @@ class AmazonConverseConfig(BaseConfig):
                 )
                 thinking["budget_tokens"] = BEDROCK_MIN_THINKING_BUDGET_TOKENS
 
+    def _model_accepts_anthropic_thinking_param(self, model: str, base_model: str) -> bool:
+        """Whether the model accepts the Anthropic-shaped ``thinking`` / ``reasoning_effort`` request field.
+
+        The Converse mapping serializes ``thinking`` into ``additionalModelRequestFields`` in Anthropic's
+        shape. Only Anthropic Claude reasoning models accept that field; DeepSeek models reason natively
+        and reject it (a 400 when it leaks through), even though they advertise ``supports_reasoning``.
+        """
+        if "deepseek" in model or "deepseek" in base_model:
+            return False
+        return (
+            "claude-3-7" in model
+            or "claude-sonnet-4" in model
+            or "claude-opus-4" in model
+            or supports_reasoning(model=model, custom_llm_provider=self.custom_llm_provider)
+            or supports_reasoning(model=base_model, custom_llm_provider=self.custom_llm_provider)
+        )
+
     def get_supported_openai_params(self, model: str) -> List[str]:
         from litellm.utils import supports_function_calling
 
@@ -553,17 +570,7 @@ class AmazonConverseConfig(BaseConfig):
             # Nova 2 models support reasoning_effort (transformed to reasoningConfig)
             # These models use a different reasoning structure than Anthropic's thinking parameter
             supported_params.append("reasoning_effort")
-        elif (
-            "claude-3-7" in model
-            or "claude-sonnet-4" in model
-            or "claude-opus-4" in model
-            or "deepseek.r1" in model
-            or supports_reasoning(
-                model=model,
-                custom_llm_provider=self.custom_llm_provider,
-            )
-            or supports_reasoning(model=base_model, custom_llm_provider=self.custom_llm_provider)
-        ):
+        elif self._model_accepts_anthropic_thinking_param(model=model, base_model=base_model):
             supported_params.append("thinking")
             supported_params.append("reasoning_effort")
 
diff --git a/tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py b/tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py
index fc12ead36a1..f7e2716bd19 100644
--- a/tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py
+++ b/tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py
@@ -553,6 +553,58 @@ def test_get_supported_openai_params():
     assert "reasoning_effort" in supported_params
 
 
+@pytest.mark.parametrize(
+    "model",
+    [
+        "bedrock/us.deepseek.r1-v1:0",
+        "bedrock/converse/us.deepseek.r1-v1:0",
+        "bedrock/deepseek.v3-v1:0",
+        "bedrock/deepseek.v3.2",
+    ],
+)
+def test_bedrock_deepseek_does_not_advertise_thinking(model):
+    """DeepSeek reasons natively on Bedrock and rejects the Anthropic-shaped
+    `thinking`/`reasoning_effort` field, so it must not be advertised as supported
+    (otherwise it leaks into additionalModelRequestFields and Bedrock 400s)."""
+    config = AmazonConverseConfig()
+    supported_params = config.get_supported_openai_params(model=model)
+    assert "thinking" not in supported_params
+    assert "reasoning_effort" not in supported_params
+
+
+def test_bedrock_deepseek_r1_thinking_raises_without_drop_params():
+    """Passing `thinking` to Bedrock DeepSeek R1 must fail client-side with a clear
+    UnsupportedParamsError instead of leaking through and hitting a Bedrock 400."""
+    with pytest.raises(litellm.UnsupportedParamsError):
+        litellm.utils.get_optional_params(
+            model="us.deepseek.r1-v1:0",
+            custom_llm_provider="bedrock",
+            thinking={"type": "enabled", "budget_tokens": 1024},
+        )
+
+
+def test_bedrock_deepseek_r1_thinking_dropped_does_not_leak_into_request():
+    """With drop_params, `thinking` is dropped rather than forwarded into
+    additionalModelRequestFields for Bedrock DeepSeek R1."""
+    optional_params = litellm.utils.get_optional_params(
+        model="us.deepseek.r1-v1:0",
+        custom_llm_provider="bedrock",
+        thinking={"type": "enabled", "budget_tokens": 1024},
+        drop_params=True,
+    )
+    assert "thinking" not in optional_params
+
+    config = AmazonConverseConfig()
+    request = config._transform_request(
+        model="bedrock/converse/us.deepseek.r1-v1:0",
+        messages=[{"role": "user", "content": "Say hi in one word."}],
+        optional_params=optional_params,
+        litellm_params={},
+        headers={},
+    )
+    assert "thinking" not in request.get("additionalModelRequestFields", {})
+
+
 def test_get_supported_openai_params_bedrock_converse():
     """
     Test that all documented bedrock converse models have the same set of supported openai params when using