Add NanoV3 reasoning parser support (#15113)

python/sglang/srt/entrypoints/openai/serving_chat.py

@@ -1065,8 +1065,8 @@ class OpenAIServingChat(OpenAIServingBase):
                 request.chat_template_kwargs is not None
                 and request.chat_template_kwargs.get("thinking") is True
             )
-        if self.reasoning_parser in ["qwen3", "glm45"]:
+        if self.reasoning_parser in ["qwen3", "glm45", "nano_v3"]:
-            # qwen3 and glm45 are reasoning by default
+            # qwen3, glm45, and nano_v3 are reasoning by default
             return (
                 not request.chat_template_kwargs
                 or request.chat_template_kwargs.get("enable_thinking", True) is True

python/sglang/srt/parser/reasoning_parser.py

@@ -274,6 +274,22 @@ class MiniMaxAppendThinkDetector(BaseReasoningFormatDetector):
         return StreamingParseResult(normal_text=self.think_start_token + text)
 
 
+class NanoV3Detector(BaseReasoningFormatDetector):
+    """
+    Detector for NanoV3 model.
+    Uses the same reasoning format as DeepSeek-R1: (<think>)*(.*)</think>
+
+    """
+
+    def __init__(self, stream_reasoning: bool = True, force_reasoning: bool = False):
+        super().__init__(
+            "<think>",
+            "</think>",
+            force_reasoning=force_reasoning,
+            stream_reasoning=stream_reasoning,
+        )
+
+
 class ReasoningParser:
     """
     Parser that handles both streaming and non-streaming scenarios for extracting
@@ -297,6 +313,7 @@ class ReasoningParser:
         "minimax": Qwen3Detector,
         "minimax-append-think": MiniMaxAppendThinkDetector,
         "step3": DeepSeekR1Detector,
+        "nano_v3": NanoV3Detector,
     }
 
     def __init__(