I see empty <think> tags in the history when using GLM 4.5 Air, is that on purpose?

[Madd0g]

Does it make sense that reasoning is stripped from history? I see only empty tags in the the logs even though I saw the LLM responding with reasoning... is it stripped on purpose on following conversation turns?

[awni]

It is stripped in some cases. What command are you using?

[Madd0g]

I've been trying out the mlx-omni-server.

In the server logs, I see <think></think>

[awni]

I would ask your question in mlx-omni-server then as I am not sure how they hanlde thinking traces. It is definitely possible they get stripped.

[Madd0g]

I looked at the inference code (how it uses mlx-lm) and didn't see anything. I guess I assumed it's something the embedded template for the model might control, was hoping for a pointer.

[awni]

That model is a hybrid reasoning model. You can control whether or not it will go into thinking mode by sending the right arguments to the chat template (e.g. enable_thinking) or putting \nothink in the prompt.

I'm not sure that's what's happening here since it sounds like you were seeing the thinking traces in the log and then they got removed in the response.