Incorrect thinking response in self-hosted deepseek-r1-distill-qwen-32b

[MateuszKepczynski]

What Happened?

When making a request to self-hosted model deepseek-r1-distill-qwen-32b to /v1/chat/completions, model thinking is incorrect formatted or provided.

curl:

curl --request POST \
  --url http://localhost:8787/v1/chat/completions \
  --header 'authorization: ****' \
  --header 'content-type: application/json' \
  --header 'x-portkey-custom-host: http://localhost:8080/openai/v1' \
  --header 'x-portkey-provider: openai' \
  --data '{
  "max_tokens": 1024,
  "messages": [
    {
      "role": "user",
      "content": "hi"
    }
  ],
  "model": "deepseek-r1-distill-qwen-32b"
}'

Response:

{
  "id": "*****",
  "object": "chat.completion",
  "created": 1757070852,
  "model": "deepseek-r1-distill-qwen-32b",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "reasoning_content": null,
        "content": "Okay, so the user greeted me with \"hi\". I need to respond in a friendly and welcoming manner.\n</think>\n\nHello! 👋 How can I assist you today?",
        "tool_calls": []
      },
      "logprobs": null,
      "finish_reason": "stop",
      "stop_reason": null
    }
  ],
  "usage": {
    "prompt_tokens": 81,
    "total_tokens": 252,
    "completion_tokens": 171,
    "prompt_tokens_details": null
  },
  "prompt_logprobs": null
}

As you can see, the response "content" only contains a closure tag - </think>.

What Should Have Happened?

Opening of the HTML <think> tag should be included in content or thinking should be found in reasoning_content object.

Relevant Code Snippet

No response

Your Twitter/LinkedIn

No response