How to Disable Thinking Output in Qwen3.5 Models Using LM Studio

Qwen3.5 lightweight models lack the Think soft switch found in larger Qwen‑32B models, which can be inconvenient. The following steps show how to disable the thinking output in LM Studio by creating a virtual model directory and configuring model.yaml.

1. Quick Summary (What to Do)

Keep the original base model directory unchanged.

Create a new virtual model directory named -no-thinking.

Write a correct model.yaml file in that directory.

Load the new virtual model in LM Studio.

Open a fresh chat window to verify the setting.

Key principle: Do not place the virtual model directly inside the base model directory.

2. Directory Structure (Correct Layout)

Assuming the base model is already downloaded:

Base directory: /Users/zz/.lmstudio/models/unsloth/Qwen3.5-27B-GGUF File: Qwen3.5-27B-Q2_K.gguf Create a new directory alongside it:

/Users/zz/.lmstudio/models/unsloth/
├── Qwen3.5-27B-GGUF/
│   ├── Qwen3.5-27B-Q2_K.gguf
│   └── mmproj-F32.gguf
└── Qwen3.5-27B-GGUF-no-thinking/
    └── model.yaml

3. Ready‑to‑Use model.yaml

Save the following content to

/Users/zz/.lmstudio/models/unsloth/Qwen3.5-27B-GGUF-no-thinking/model.yaml

:

model: unsloth/Qwen3.5-27B-GGUF-no-thinking
base: unsloth/Qwen3.5-27B-GGUF/Qwen3.5-27B-Q2_K.gguf
metadataOverrides:
  reasoning: false
customFields:
  - key: enableThinking
    displayName: "Enable Thinking"
    description: "Whether to allow thinking output before the final answer"
    type: boolean
    defaultValue: false
    effects:
      - type: setJinjaVariable
        variable: enable_thinking

The core points of this configuration are: defaultValue: false – disables thinking by default. variable: enable_thinking – the variable name must be exact.

4. Operations in LM Studio

Open LM Studio and refresh the model list.

Select and load unsloth/Qwen3.5-27B-GGUF-no-thinking.

In the model configuration, confirm that Enable Thinking = Off.

Open a new chat window and test.

Suggested test prompt: 用一句话介绍你自己。 If successful, the model returns the final answer directly without a preceding reasoning block.

5. Why Does Thinking Still Appear?

The three most common reasons are:

1) Old session overrides the model default

An existing chat may retain enableThinking=true at the session level.

Solution: start a new chat window, or manually verify the Enable Thinking setting in the current session.

2) Variable name typo

Common mistake: disable_thinking. The correct name is enable_thinking.

3) Loading the wrong model

Make sure to load the virtual model unsloth/Qwen3.5-27B-GGUF-no-thinking instead of the base model directory unsloth/Qwen3.5-27B-GGUF/….

6. Common Error Quick‑Reference

Invalid file name (repo field)

Cause: repo contains a sub‑path.

Fix: Use only the repository name without any /xxx suffix.

Circular dependency detected

Cause: Virtual model and base model share the same ID, creating a self‑reference.

Fix: Assign a distinct model ID to the virtual model.

A dependent model is missing

Cause: Target dependency does not exist or the base directory was mistaken for the virtual model.

Fix: Keep the base model directory unchanged and create the virtual model directory separately.

Expected array, received object (config.operation.fields)

Cause: Field type mismatch.

Fix: Remove fields that do not conform to the schema until the manifest validates.

Failed to load model (Exit code: null)

Cause: Usually not a YAML issue but insufficient resources (GPU/Metal OOM).

Fix: Reduce context length, lower concurrency, or decrease GPU offload.

7. Practical Tips

Start with the minimal model.yaml configuration, verify it works, then add advanced settings.

Reload the model after each configuration change.

When troubleshooting “still thinking”, prioritize testing in a new session to avoid leftover state.

8. One‑Sentence Summary

In LM Studio, the most reliable way to turn off thinking is to keep the original model directory untouched, create a separate -no-thinking virtual model, and ensure the enable_thinking variable maps correctly with a default value of false .