From 3c34d2c3ebfe0d4dfc02353980ed50a350fa0324 Mon Sep 17 00:00:00 2001
From: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>
Date: Wed, 28 Jan 2026 22:47:09 -0500
Subject: [PATCH] [FIX] kimi_k2 reasoning parser (#17901)

Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
---
 .../srt/entrypoints/openai/serving_chat.py    | 15 +++++++++----
 python/sglang/srt/parser/reasoning_parser.py  | 22 +------------------
 2 files changed, 12 insertions(+), 25 deletions(-)

diff --git a/python/sglang/srt/entrypoints/openai/serving_chat.py b/python/sglang/srt/entrypoints/openai/serving_chat.py
index f833120db..ab4d6c7fb 100644
--- a/python/sglang/srt/entrypoints/openai/serving_chat.py
+++ b/python/sglang/srt/entrypoints/openai/serving_chat.py
@@ -1196,16 +1196,23 @@ class OpenAIServingChat(OpenAIServingBase):
         """Judge whether the request needs reasoning"""
         if not self.reasoning_parser:
             return False
-        if self.reasoning_parser in ["deepseek-v3", "kimi_k2"]:
+        if self.reasoning_parser in ["deepseek-v3"]:
+            # Models that require explicit enable thinking (thinking=True)
             return (
                 request.chat_template_kwargs is not None
                 and request.chat_template_kwargs.get("thinking") is True
             )
-        if self.reasoning_parser in ["qwen3", "glm45", "nano_v3", "interns1"]:
-            # qwen3, glm45, nano_v3, and interns1 are reasoning by default
+        if self.reasoning_parser in ["kimi_k2"]:
+            # Models that thinking by default, and can be disabled by setting thinking=False
             return (
                 not request.chat_template_kwargs
-                or request.chat_template_kwargs.get("enable_thinking", True) is True
+                or request.chat_template_kwargs.get("thinking") is not False
+            )
+        if self.reasoning_parser in ["qwen3", "glm45", "nano_v3", "interns1"]:
+            # Models that thinking by default, and can be disabled by setting enable_thinking=False
+            return (
+                not request.chat_template_kwargs
+                or request.chat_template_kwargs.get("enable_thinking") is not False
             )
         return True  # default
 
diff --git a/python/sglang/srt/parser/reasoning_parser.py b/python/sglang/srt/parser/reasoning_parser.py
index 87535d95e..455101b5c 100644
--- a/python/sglang/srt/parser/reasoning_parser.py
+++ b/python/sglang/srt/parser/reasoning_parser.py
@@ -179,26 +179,6 @@ class DeepSeekR1Detector(BaseReasoningFormatDetector):
         # https://github.com/sgl-project/sglang/pull/3202#discussion_r1950153599
 
 
-class KimiK2Detector(BaseReasoningFormatDetector):
-    """
-    Detector for Kimi K2 model.
-
-    It uses the DeepSeek-R1 reasoning format: (<think>)*(.*)</think>.
-    Defaults to thinking mode (force_reasoning=True), but allows disabling it
-    if the model is configured to not think.
-    """
-
-    def __init__(self, stream_reasoning: bool = True, force_reasoning: bool = True):
-        super().__init__(
-            "<think>",
-            "</think>",
-            # Allow force_reasoning to be controlled by arguments, defaulting to True
-            # to match vLLM's default `thinking=True` behavior.
-            force_reasoning=force_reasoning,
-            stream_reasoning=stream_reasoning,
-        )
-
-
 class Qwen3Detector(BaseReasoningFormatDetector):
     """
     Detector for Qwen3 models (e.g., Qwen/Qwen3-235B-A22B).
@@ -398,7 +378,7 @@ class ReasoningParser:
         "glm45": Qwen3Detector,
         "gpt-oss": GptOssDetector,
         "kimi": KimiDetector,
-        "kimi_k2": KimiK2Detector,
+        "kimi_k2": Qwen3Detector,
         "qwen3": Qwen3Detector,
         "qwen3-thinking": Qwen3Detector,
         "minimax": Qwen3Detector,