Force disable thinking for gemini-2.5-flash model?

[Dima-Mediator]

Following #10121

Does anybody understand how to disable thinking for the gemini-2.5-flash-preview-04-17 model on LiteLLM side?

The client (Open WebUI) is unfortunately not currently capable of transmitting the Anthropic-style "thinking" parameter and only supports reasoning_effort which is not enough. And thinking is currently enabled by default for this model.

In the original PR, someone suggested using using YAML model-level litellm_params to force a thinking setting (with budget 0) but I have not figured out the exact syntax to make it work. Has someone succeeded?

[gaambo]

thinking={"type": "disabled", "budget_tokens": 0} seems to work for me (at least I think?).
Although "type": "disabled" is probably useless here, it's about the budget tokens.

[Dima-Mediator]

Sorry, can you clarify where exactly you have put this in? API Request, or LiteLLM YAML config?

If we control our own API request then yes I could add that to the request.

Problem is that I have not found a way to add this parameter on the client side (Open WebUI)

[gaambo]

I'm using that as an argument to the completion function directly in Python.

completion(
                model=model,
                messages=[{"role": "system", "content": system_prompt}, {"role": "user", "content": prompt}],
                thinking={"type": "disabled", "budget_tokens": 0},
            )

Sorry, I missed the part about using Open WebUI in the OT and just wanted to share my solution.