Clarification Needed: Does chat-model internally use completion-model during chat responses?

[daoutech-sujinlee]

I'm currently using Tabby with both a completion-model and a chat-model configured, like this:
tabby serve --model DeepSeekCoder-1.3B --chat-model Qwen2.5-Coder-7B-Instruct

According to the documentation, the completion model and the chat model operate independently — the completion model is for code suggestions in editors, and the chat model is for conversational queries.

However, during my tests, I observed significantly different chat results depending on which completion model was paired with the same chat model. This led me to wonder:

Does Tabby internally utilize the completion-model in the background to assist chat-model responses?

Or are they completely separate as the docs suggest?

It would be great to have an official clarification on how the models interact during chat-based usage, and whether the completion model has any implicit influence on chat output.

In my case, I tested the following combinations:

This makes it seem like the completion model might be influencing the chat result, even though the chat model is the same.

I'd appreciate any clarification from the team — and if this is the intended behavior, it would be helpful to include in the documentation.
To clarify, all of my tests were done via the Tabby chat feature within the IDE (IntelliJ).

[zwpaper]

Hello @daoutech-sujinlee, Tabby does use the two models independently.

Based on your post, the issue you're experiencing could be related to the small chat model you're using; it's possible to receive different answers upon retrying the questions, even without any changes being made.