Feature Request: WebUI add reasoning effort level on a per-message basis

[engrtipusultan]

Prerequisites

• I am running the latest code. Mention the version if possible as well.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new and useful enhancement to share.

Feature Description

I would like to propose a feature enhancement for our WebUI: the ability to dynamically change a model's reasoning effort level on a per-message basis at runtime.

Currently, for models like GPT OSS that support it, we can pass a reasoning_effort parameter (e.g., low, medium, high) via kwargs or directly in the API request.

Proposed Solution:
Add a dedicated button or dropdown selector within the message input pane. This would allow users to select the reasoning effort level (e.g., Low, Medium, High, Off, On) for each new message they send, without needing to alter the overall session configuration.

Extended Use Case & Models:
This functionality would be highly valuable for other models that support similar runtime parameters, such as: Qwen3 ,NVIDIA Nemotron toggling reasoning.

This feature would provide greater flexibility and control during conversations, enabling users to optimize for speed or depth as needed for each query.

Motivation

It is already implemented in Cherry Studio and some other clients.

Possible Implementation

No response

[pwilkin]

This is planned, but requires #17750 before it's done. Unless you mean only setting it for models that are trained with support for a reasoning_effort switch - but then I don't think any GGML-compatible model outside of GPT-OSS natively supports it.

[engrtipusultan]

If it is planned then it is great. I thought other clients have it and it will be good idea to float, if it is feasible to implement in llama.cpp frontend.

[aldehir]

If it is planned then it is great. I thought other clients have it and it will be good idea to float, if it is feasible to implement in llama.cpp frontend.

It's a bit tricky because not every model supports levels like low/medium/high. For some it's an on/off toggle, for others it's a token budget.

This feature would require some way to query the backend for model reasoning capabilities. Something we're currently exploring as well. It's a good idea, but an adequate solution requires a bit more plumbing within llama.cpp.

[engrtipusultan]

I appreciate you guys considering it. My idea was to bring it to attention, as I believe it will be useful for llama.cpp users.

In other thread someone mentioned it is holiday season for some. Enjoy your holidays if it is so, with time I am sure you guys will device whatever is deemed best.

[segmond]

Even reasoning on/off will be good.

[sultanqasim]

There are a lot of models with toggleable reasoning (Qwen, GLM, DeekSeek, etc). It would be very useful to have a reasoning toggle in the WebUI.

[sultanqasim]

Related: #20001

[NickM-27]

This would definitely be awesome, with the current thinking budget feature seems like we are getting closer

[vanmilleru]

#18405
#21141

[DoS007]

Lightweightness is imo an important argument, cause this makes model reloading unnecessary (which takes time):

E.g. Gemma 4 llama.cpp server start parameter for configuring reasoning:

• Disable reasoning: --reasoning off
• Enable reasoning: --chat-template-kwargs '{"enable_thinking":true}'

Each start takes time to load up the model. But switching the reasoning from disabled to enabled or vice versa doesn't need the model to be completely reloaded? So a switch would make that switching much more lightweight (without reloading the model).