model: support Kimi-K2.5 (#17789)

Co-authored-by: Mick <mickjagger19@icloud.com>

python/sglang/srt/parser/reasoning_parser.py

@@ -148,6 +148,26 @@ 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).
@@ -307,7 +327,7 @@ class ReasoningParser:
         "glm45": Qwen3Detector,
         "gpt-oss": GptOssDetector,
         "kimi": KimiDetector,
-        "kimi_k2": DeepSeekR1Detector,
+        "kimi_k2": KimiK2Detector,
         "qwen3": Qwen3Detector,
         "qwen3-thinking": Qwen3Detector,
         "minimax": Qwen3Detector,