From ee779d4259b19bb10f526eede7e51fc599307b65 Mon Sep 17 00:00:00 2001
From: NasonZ <ntzikayo@gmail.com>
Date: Fri, 11 Sep 2026 12:56:55 +0100
Subject: [PATCH] fix(deepseek): forward graded reasoning_effort instead of
 discarding it

map_openai_params collapsed every non-"none" reasoning_effort into a bare
thinking: {"type": "enabled"} and dropped the level, so low, high and max
were all indistinguishable from the default and callers could not ask for
reduced reasoning at all

DeepSeek accepts reasoning_effort directly in thinking mode and grades it
server-side, so the value is forwarded rather than remapped here

Fixes #27439
---
 litellm/llms/deepseek/chat/transformation.py  |  8 +++-
 .../chat/test_deepseek_chat_transformation.py | 39 +++++++++++++++++++
 2 files changed, 45 insertions(+), 2 deletions(-)

diff --git a/litellm/llms/deepseek/chat/transformation.py b/litellm/llms/deepseek/chat/transformation.py
index ea19a7c7ddf..ea0693512de 100644
--- a/litellm/llms/deepseek/chat/transformation.py
+++ b/litellm/llms/deepseek/chat/transformation.py
@@ -39,7 +39,8 @@ class DeepSeekChatConfig(OpenAIGPTConfig):
         Handles `thinking` and `reasoning_effort` parameters for DeepSeek reasoner models.
         DeepSeek supports `{"type": "enabled"}` and `{"type": "disabled"}` - no budget_tokens
         like Anthropic. `reasoning_effort="none"` is the OpenAI-style way to ask for thinking
-        off, so it maps to `{"type": "disabled"}`; any other effort keeps thinking on.
+        off, so it maps to `{"type": "disabled"}`; any other effort keeps thinking on and
+        is forwarded, since DeepSeek grades effort itself server-side.
 
         Reference: https://api-docs.deepseek.com/guides/thinking_mode
         """
@@ -57,7 +58,10 @@ class DeepSeekChatConfig(OpenAIGPTConfig):
 
         # Otherwise fall back to reasoning_effort: "none" disables, anything else enables
         elif reasoning_effort is not None:
-            optional_params["thinking"] = {"type": "disabled" if reasoning_effort == "none" else "enabled"}
+            thinking_enabled: Final = reasoning_effort != "none"
+            optional_params["thinking"] = {"type": "enabled" if thinking_enabled else "disabled"}
+            if thinking_enabled:
+                optional_params["reasoning_effort"] = reasoning_effort
 
         return optional_params
 
diff --git a/tests/test_litellm/llms/deepseek/chat/test_deepseek_chat_transformation.py b/tests/test_litellm/llms/deepseek/chat/test_deepseek_chat_transformation.py
index 3f93264d0d0..702385a277e 100644
--- a/tests/test_litellm/llms/deepseek/chat/test_deepseek_chat_transformation.py
+++ b/tests/test_litellm/llms/deepseek/chat/test_deepseek_chat_transformation.py
@@ -443,6 +443,7 @@ class TestDeepSeekThinkingParams:
         )
 
         assert result["thinking"] == {"type": "enabled"}
+        assert result["reasoning_effort"] == "medium"
 
     def test_map_reasoning_effort_low(self):
         """Test that reasoning_effort='low' maps to thinking enabled."""
@@ -457,6 +458,7 @@ class TestDeepSeekThinkingParams:
         )
 
         assert result["thinking"] == {"type": "enabled"}
+        assert result["reasoning_effort"] == "low"
 
     def test_map_reasoning_effort_high(self):
         """Test that reasoning_effort='high' maps to thinking enabled."""
@@ -471,6 +473,43 @@ class TestDeepSeekThinkingParams:
         )
 
         assert result["thinking"] == {"type": "enabled"}
+        assert result["reasoning_effort"] == "high"
+
+    def test_map_graded_effort_is_forwarded_not_collapsed(self):
+        """Every effort value reaches DeepSeek, which grades it server-side."""
+        for effort in ("minimal", "low", "medium", "default", "high", "xhigh", "max", "ultra"):
+            result = self.config.map_openai_params(
+                non_default_params={"reasoning_effort": effort},
+                optional_params={},
+                model=self.model,
+                drop_params=False,
+            )
+            assert result["thinking"] == {"type": "enabled"}
+            assert result["reasoning_effort"] == effort
+
+    def test_explicit_thinking_is_honoured_exactly_as_given(self):
+        """An explicit toggle decides thinking, and no effort is merged into it."""
+        result = self.config.map_openai_params(
+            non_default_params={"thinking": {"type": "enabled"}, "reasoning_effort": "low"},
+            optional_params={},
+            model=self.model,
+            drop_params=False,
+        )
+        assert result["thinking"] == {"type": "enabled"}
+        assert "reasoning_effort" not in result
+
+    def test_disabled_thinking_carries_no_effort(self):
+        """Nothing is thinking, so an effort value would be meaningless."""
+        for params in ({"reasoning_effort": "none"},
+                       {"thinking": {"type": "disabled"}}):
+            result = self.config.map_openai_params(
+                non_default_params=params,
+                optional_params={},
+                model=self.model,
+                drop_params=False,
+            )
+            assert result["thinking"] == {"type": "disabled"}
+            assert "reasoning_effort" not in result
 
     def test_map_reasoning_effort_none_does_not_enable_thinking(self):
         """Test that reasoning_effort='none' does not enable thinking."""