### Check for existing issues

- [x] I have searched the existing issues and checked that my issue is not a duplicate.

### What happened?

I deployed both a cloud model and a local Qwen14B model using LiteLLM. I noticed that when calling the local 14B model, the responses are always in JSON format. I tried disabling thinking mode, enabling streaming output, and modifying the config, but none of these changes worked.

Can anyone help explain how to properly solve this issue?

### Steps to Reproduce

* Changed reasoning: true → Probably unrelated, because other models are also running with false and work fine. * Added a system_prompt telling the model not to output JSON → Ineffective guess. * Enabled stream: true → Ineffective guess. * Modified SOUL.md → The problem was not related to this. * Created qwen3-nothink to remove thinking tags → Direction was partially correct, but repeated attempts caused errors and created duplicate models, wasting disk space. * Rebuilt the Modelfile multiple times → Re-downloaded the 9GB model every time. * Modified feishu-channel-rules → Problem was unrelated. Repeatedly deleted and recreated models based on guesses …