Handling of models pulled from HF does not properly handle thinking

[Expro]

LocalAI version:
Any, including latest commit as on now.

Describe the bug
When user pulls thinking model from HF, default model configuration file does not handle "thinking" part correctly. Token stream outputted by model does not properly tag thinking part, making it appear as part of naswer.

To Reproduce
Pull any thinking model from HF, chat with it, observe no distinction between thinking and providing final answer.

[Expro]

@mudler I have tested version with relevant commit, it still does not work and emits thinking content as part of answer.

[mudler]

@mudler I have tested version with relevant commit, it still does not work and emits thinking content as part of answer.

I've tested with qwen 3 thinking and it works. I didn't have a look at nemotron closely, suspecting it might be different thinking tags. What does return via API?

[Expro]

Doesn't work for me even on Qwen3 Thinking:

Commit 2de3044, so no, it's not old version.

[Expro]

GLM-4.7-Flash - doesn't work as well. Overall - it's doesn't work.

[mudler]

Doesn't work for me even on Qwen3 Thinking:

The ones from the gallery works, the problem is only with the imported models. This #8132 should fix for the models being imported