diff --git a/xinference/model/llm/tests/test_utils.py b/xinference/model/llm/tests/test_utils.py
index de445cf9ca..3691b8ebda 100644
--- a/xinference/model/llm/tests/test_utils.py
+++ b/xinference/model/llm/tests/test_utils.py
@@ -2830,6 +2830,46 @@ def test_post_process_completion_with_parser():
     ), f"Mismatch: expected {expected_filtered}, got {result_filtered}"
 
 
+def test_post_process_completion_stopped_while_thinking():
+    # Generation hits max_tokens inside the thinking block, so the output has
+    # no tool call and no content. Like the request without tools, the
+    # response must keep "length" and must not repeat the reasoning as content.
+    mixin = ChatModelMixin()
+    mixin.tool_parser = QwenToolParser()
+    mixin.reasoning_parser = ReasoningParser(
+        reasoning_content=True,
+        reasoning_start_tag="<think>",
+        reasoning_end_tag="</think>",
+        enable_thinking=True,
+    )
+    test_case = {
+        "id": "1",
+        "object": "text_completion",
+        "created": 0,
+        "model": "qwen3",
+        "choices": [
+            {
+                "text": "The user asks for the weather in Shanghai, so I should call",
+                "index": 0,
+                "logprobs": None,
+                "finish_reason": "length",
+            }
+        ],
+        "usage": {"prompt_tokens": 159, "completion_tokens": 12, "total_tokens": 171},
+    }
+
+    result = mixin._post_process_completion(None, model_uid="qwen3", c=test_case)
+
+    choice = result["choices"][0]
+    assert choice["finish_reason"] == "length"
+    assert choice["message"]["content"] == ""
+    assert choice["message"]["tool_calls"] == []
+    assert (
+        choice["message"]["reasoning_content"]
+        == "The user asks for the weather in Shanghai, so I should call"
+    )
+
+
 # ── Security tests for Llama3ToolParser ────────────────────
 
 
diff --git a/xinference/model/llm/utils.py b/xinference/model/llm/utils.py
index 784499035a..e91d4794f1 100644
--- a/xinference/model/llm/utils.py
+++ b/xinference/model/llm/utils.py
@@ -1155,9 +1155,10 @@ def _post_process_completion(
             reasoning_content, processed_content = (
                 self.reasoning_parser.extract_reasoning_content(text)
             )
-            # Use the processed content (without thinking tags) for tool parsing
-            if processed_content:
-                text = processed_content
+            # Use the processed content (without thinking tags) for tool parsing.
+            # It is empty when the output only contains reasoning, e.g. when
+            # generation stops inside the thinking block.
+            text = processed_content or ""
 
         # Then, extract tool calls from the processed text (without thinking tags)
         tool_calls = []
@@ -1192,7 +1193,11 @@ def _post_process_completion(
             # For non-tool calls, use the processed content from reasoning parser
             content = text
 
-        finish_reason = "tool_calls" if tool_calls else "stop"
+        finish_reason = (
+            "tool_calls"
+            if tool_calls
+            else c["choices"][0].get("finish_reason") or "stop"
+        )
 
         m = {
             "role": "assistant",