Skip to content

<think> tags inserted for Qwen3.5 0.8B regardless of the reasoning toggle #401

@hlibr

Description

@hlibr

Describe the bug

When using the Qwen3.5 0.8B model, the following behavior is observed:

Regardless of the reasoning toggle state (on or off) in the LLM component, the empty think block always gets appended to the prompt:

Image Image

Qwen3.5 0.8B supports thinking mode, however it is off by default, and can only be enabled by adding "enable_thinking": True to the request's extra_body.

The think block thus does nothing except potentially degrade the performance.

Backtested with Qwen3 0.6B - with that model the think block only gets appended with reasoning off.

LLMUnity version

3.0.3

Operating System

macOs

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions