[Bug] Qwen3.5-9B (GGUF): enableThinking=false via API ignored — reasoning_content consumes all max_tokens, content empty

[Deepwater1000]

Environment

• LM Studio: 0.4.15+2
• OS: macOS (Apple Silicon M4)
• Model: qwen3.5-9b (Q4_K_M GGUF, 5.2GB)
• Backend: llama.cpp

Bug Description

When sending a simple text-generation request to the local OpenAI-compatible API (/v1/chat/completions), enableThinking: false is completely ignored. The model still outputs reasoning_content (thinking), and when max_tokens is limited, the thinking process consumes all tokens, leaving message.content empty.

This breaks any downstream consumer that reads message.content for non-structured-output use cases (e.g., summarization, compression, classification).

Reproduction

1. enableThinking: false via API — FAILS

curl -s -X POST http://localhost:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.5-9b",
    "messages": [
      {"role": "system", "content": "直接输出结果，不要思考，简洁回答。"},
      {"role": "user", "content": "1+1等于几？"}
    ],
    "max_tokens": 100,
    "temperature": 0.1,
    "enableThinking": false
  }'

Response:

{
  "choices": [{
    "message": {
      "content": "",
      "reasoning_content": "Thinking Process:\n1. Analyze the Request..."
    }
  }],
  "usage": {
    "completion_tokens": 100,
    "completion_tokens_details": {
      "reasoning_tokens": 99
    }
  }
}

• content = empty
• reasoning_tokens = 99 out of 100 tokens
• The only reason content is empty is that max_tokens ran out during thinking

2. chat_template_kwargs workaround — ALSO FAILS

Following the workaround from #1559:

curl -s -X POST http://localhost:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.5-9b",
    "messages": [
      {"role": "user", "content": "写一句关于春天的诗"}
    ],
    "max_tokens": 100,
    "temperature": 0.1,
    "chat_template_kwargs": {"enable_thinking": false}
  }'

Result: Same behavior — reasoning_content still present, content empty.

3. defaultValue: false in model.yaml — ALSO FAILS

Edited ~/.lmstudio/hub/models/qwen/qwen3.5-9b/model.yaml:

customFields:
  - key: enableThinking
    defaultValue: false  # ← explicitly set to false

Then reloaded the model. API behavior unchanged — thinking still fires.

Root Cause Analysis

Even with all three layers of suppression (API parameter, chat_template_kwargs, defaultValue in model.yaml), the Qwen3.5 model still generates a \<think>...\ block and outputs it as reasoning_content in the API response.

The key distinction from existing issues:

• Unable to close QWen3's thinks mode in LMStudio. #632 and Qwen 3.5 Model: Switching Between Thinking and Non-Thinking #1559 discuss the general inability to disable thinking in the UI / API
• Structured output produces an empty content, and the content goes to reasoning_content in the response #1698, response_format: json_schema applies schema constraint to reasoning stream on Qwen 3.5, leaving content empty #1773, [Bug]: qwen3.5-4b-mlx structured output returns JSON in reasoning_content while content is empty #1971 discuss structured output (response_format: json_schema) where JSON ends up in reasoning_content

Our bug is different: This is plain text generation (no response_format). The symptom is:

When max_tokens is set, thinking eats the entire token budget, leaving content empty.

This is catastrophic for automated pipelines that rely on message.content — summarization scripts, classification pipelines, compression agents, etc.

Expected Behavior

When enableThinking: false is set (by any means — API param, chat_template_kwargs, or model.yaml defaultValue), the model should:

1. Not produce any <think> blocks or reasoning_content
2. Output the final response directly in message.content

If suppressing thinking is technically impossible for Qwen3.5, then at minimum LM Studio should:

• Honor a token budget split: reserve some tokens for content even if thinking overflows
• Or document clearly that enableThinking: false has no effect on Qwen3.5

Workaround Found by Community

Changing chat template to ChatML in LM Studio Developer → Inference → Prompt Template suppresses thinking, but this is a blunt workaround that may break vision/tool-use capabilities.

[jtarchie]

I have also experienced this with gemma v4 models too.