<think> block missing in API response (LM Studio 0.3.22, Qwen3-Thinking models)

[burdzi]

Which version of LM Studio?
LM Studio 0.3.22-1 (Linux)

Which operating system?
Ubuntu 24.04.2 LTS

What is the bug?
When calling the HTTP API (/v1/chat/completions) with a "Thinking" variant of Qwen3 (e.g., qwen3-4b-thinking-2507), the <think>...</think> block is missing from the response.

This used to work correctly in previous versions of LM Studio — the block was included in the API response and could be parsed by my local application. After updating to 0.3.22, the content is no longer included in the API output, even though:

1. The LM Studio GUI still shows the internal Thinking blocks, indicating the model is generating them.
2. I did not change the Jinja template or any model settings.
3. Resetting all LM Studio settings had no effect.
4. The issue occurs with multiple Qwen3-thinking models, not just one.

To Reproduce

curl http://localhost:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-4b-thinking-2507",
    "messages": [
      { "role": "system", "content": "Always answer in rhymes. Today is Thursday" },
      { "role": "user", "content": "What day is it today?" }
    ],
    "temperature": 0.7,
    "max_tokens": -1,
    "stream": false
}'

Example response:
[qwen3-4b-thinking-2507] Generated prediction:

{
  "id": "chatcmpl-evf5zdsvuaonsed2l41gw",
  "object": "chat.completion",
  "created": 1754495472,
  "model": "qwen3-4b-thinking-2507",
  "choices": [
    {
      "index": 0,
      "logprobs": null,
      "finish_reason": "stop",
      "message": {
        "role": "assistant",
        "content": "\nOkay, the user asked \"What day is it today?\" but also specified that I should always respond in rhymes and that today is Thursday. \n\nHmm, this is interesting—they're testing if I follow instructions creatively while staying accurate. They probably want something playful but precise since they mentioned Thursdays specifically. \n\nI recall they emphasized \"always answer in rhymes,\" so no dry facts allowed. Gotta craft a short, cheerful couplet that confirms Thursday without sounding robotic. \n\n*Brainstorming lines...* \nFirst line should state the day clearly: \"Today's the fourth day of the week\" → too wordy? Wait, \"Thursday\" is simpler! \nSecond line needs rhyme and lightness... maybe add why it matters (\"best for plans\" feels universally relatable). \n\nUser didn't give context about their mood, but since they're asking for rhymes, they likely want fun. No need to overcomplicate—just a tiny cheerful nudge. \n\n*Double-checking*: \n- Rhyme? \"Day/plan\" works (slant rhyme, but friendly enough). \n- Thursday confirmed? Yep. \n- Short? Under 10 words per line → good. \n\n...And no markdown! Pure poetry in their request. *Nods*. Done.\n\n\nToday's the day that starts with \"Th\" and ends with \"day\"—  \n**Thursday**, bright and full of plans to make it play! 🌟"
      }
    }
  ],
  "usage": {
    "prompt_tokens": 28,
    "completion_tokens": 295,
    "total_tokens": 323
  },
  "stats": {},
  "system_fingerprint": "qwen3-4b-thinking-2507"
}

Expected behavior:
The response should include a think-block with internal reasoning (as it did in earlier versions).

[Peter-De-Ath]

This is still an issue with: LM Studio 0.3.22-2

[mattjcly]

Will be fixed in next release!

[greg-gav]

Same issue with gpt-oss-120b - no separator between reasoning block and actual output

[owlcoder13]

The same issue occurs with Qwen3-32B. I can’t find information on how to downgrade LM Studio.

[zhzLuke96]

Same issue with gpt-oss-20b

missing <|channel|> <|message|> <|end|> <|start|> all special tokens

[marquesVF]

In addition to the missing block described above, the block is also absent when reasoning_content is enabled if the request contains a tools attribute. This happens even when the model generates internal reasoning. Unless the model actually makes a tool call, in which case the reasoning is included. This suggests that the mere presence of tools suppresses the output unless a tool call occurs.

[djouallah]

Any updates?

[vova616]

this is still happening

[nvdg]

Still happens, having the same issue now with GLM-4.7-Flash

[Sam-Horry]

Still happening with GLM-4.7-Flash, tried out multiple different quants, but it still doesn't properly add to the start of the response. Using with open webui means that it interprets the whole response as the non-thinking response. Can confirm that it doesn't happen using a base install of llama.cpp using the same model downloads, so its definitely not a model or chat template issue.

[LisaMacintosh]

Having this issue as well with MiniMax 2.5

[mattjcly]

Thank you all for reports. We're investigating to fix

[LisaMacintosh]

Is this part of the 0.4.3 update or is the fix still being worked on?

[mattjcly]

0.4.4 will have the fix! Please re-open if you observe any issues with that release, or if this comes back up. Thank you everyone.