LM Studio 0.4.16 (and older)
Windows 10
RTX 3060
No "Think" toggle for Unsloth's gemma-4-12b-it GGUF, while the lmstudio-community / Google edition of the same model shows one.
Steps to reproduce
- Download unsloth/gemma-4-12b-it-GGUF
- Load the model
- Open a chat and look for the per-message "Think" toggle
Expected behavior
A "Think" toggle to enable/disable thinking on a per-message basis, just like the "standard" (lmstudio-community / Google) edition of the same model has.
Actual behavior
No "Think" toggle. Thinking is permanently disabled from the UI.
Additional context
The capability is present in the file — only the UI won't expose it.
Proof the model can think: loading the same Unsloth GGUF in llama-server (the llama.cpp server LM Studio is built on) and sending chat_template_kwargs: {"enable_thinking": true} to /v1/chat/completions makes the model think and return a reasoning_content block. Both the Unsloth and lmstudio-community GGUFs ship an embedded chat template that supports the enable_thinking variable (the templates aren't identical, but both handle it). So the thinking capability lives in the file itself.
Likely root cause: the "Think" toggle is sourced from the hub catalog model.yaml (customFields: enableThinking), which is bound via base.sources only to lmstudio-community repos — never to unsloth. So the standard edition gets the toggle and the Unsloth GGUF doesn't, even though both share the same embedded-template support for enable_thinking.
LM Studio 0.4.16 (and older)
Windows 10
RTX 3060
No "Think" toggle for Unsloth's gemma-4-12b-it GGUF, while the lmstudio-community / Google edition of the same model shows one.
Steps to reproduce
Expected behavior
A "Think" toggle to enable/disable thinking on a per-message basis, just like the "standard" (lmstudio-community / Google) edition of the same model has.
Actual behavior
No "Think" toggle. Thinking is permanently disabled from the UI.
Additional context
The capability is present in the file — only the UI won't expose it.
Proof the model can think: loading the same Unsloth GGUF in llama-server (the llama.cpp server LM Studio is built on) and sending chat_template_kwargs: {"enable_thinking": true} to /v1/chat/completions makes the model think and return a reasoning_content block. Both the Unsloth and lmstudio-community GGUFs ship an embedded chat template that supports the enable_thinking variable (the templates aren't identical, but both handle it). So the thinking capability lives in the file itself.
Likely root cause: the "Think" toggle is sourced from the hub catalog model.yaml (customFields: enableThinking), which is bound via base.sources only to lmstudio-community repos — never to unsloth. So the standard edition gets the toggle and the Unsloth GGUF doesn't, even though both share the same embedded-template support for enable_thinking.