I have searched the existing issues and discussions.
Problem Description
The new Magistral model seems to want it's own reasoning to stay in context, even though with most other reasoning models you should only keep the final answer. I haven't found a setting in Open-WebUI to change that behavior as it always removes the part inside the tags from the conversation.
Desired Solution you'd like
Add a toggle to choose whether the reasoning should be kept or excluded from the conversation.
Alternatives Considered
No response
Additional Context
No response
Originally created by @Tureti on GitHub (Jun 13, 2025).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/14963
### Check Existing Issues
- [x] I have searched the existing issues and discussions.
### Problem Description
The new Magistral model seems to want it's own reasoning to stay in context, even though with most other reasoning models you should only keep the final answer. I haven't found a setting in Open-WebUI to change that behavior as it always removes the part inside the <think> tags from the conversation.
### Desired Solution you'd like
Add a toggle to choose whether the reasoning should be kept or excluded from the conversation.
### Alternatives Considered
_No response_
### Additional Context
_No response_
@Tureti Is this why the model goes into an infinite loop? I can't get it to work with OWUI.
I keeps "thinking" about the same thing in an endless loop.
<!-- gh-comment-id:2972087846 -->
@gaby commented on GitHub (Jun 14, 2025):
@Tureti Is this why the model goes into an infinite loop? I can't get it to work with OWUI.
I keeps "thinking" about the same thing in an endless loop.
@gaby I don't think that this has anything to do with the looping. If you are using a quantized version of the small model maybe this helps you. It's from the Qwen team about their own reasoning models, but might help:
We recommend setting presence_penalty to 1.5 for quantized models to suppress repetitive outputs. You can adjust the presence_penalty parameter between 0 and 2. A higher value may occasionally lead to language mixing and a slight reduction in model performance.
<!-- gh-comment-id:2972764119 -->
@Tureti commented on GitHub (Jun 14, 2025):
@gaby I don't think that this has anything to do with the looping. If you are using a quantized version of the small model maybe this helps you. It's from the Qwen team about their own reasoning models, but might help:
> We recommend setting presence_penalty to 1.5 for quantized models to suppress repetitive outputs. You can adjust the presence_penalty parameter between 0 and 2. A higher value may occasionally lead to language mixing and a slight reduction in model performance.
<!-- gh-comment-id:2972765422 -->
@gaby commented on GitHub (Jun 14, 2025):
@Tureti I'm not using the GGUF version. I'm serving the model using the official instructions from `Mistral` with `vLLM`.
```
vllm serve mistralai/Magistral-Small-2506 --tokenizer_mode mistral --config_format mistral --load_format mistral --tool-call-parser mistral --enable-auto-tool-choice --tensor-parallel-size 2
```
When prompting it from `Open-WebUI` it goes into an infinite loop thinking.
---
EDIT: It might be a `vLLM` issue, this person reported the same issue https://huggingface.co/mistralai/Magistral-Small-2506/discussions/15
@gaby Okay. As long as it does this in the first message as well (first one is unaffected by this open-webui behavior), it has nothing to do with this issue. If it only happens after multiple turns then it might be affected, but I'm honestly not sure if it could cause looping.
<!-- gh-comment-id:2972776247 -->
@Tureti commented on GitHub (Jun 14, 2025):
@gaby Okay. As long as it does this in the first message as well (first one is unaffected by this open-webui behavior), it has nothing to do with this issue. If it only happens after multiple turns then it might be affected, but I'm honestly not sure if it could cause looping.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Originally created by @Tureti on GitHub (Jun 13, 2025).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/14963
Check Existing Issues
Problem Description
The new Magistral model seems to want it's own reasoning to stay in context, even though with most other reasoning models you should only keep the final answer. I haven't found a setting in Open-WebUI to change that behavior as it always removes the part inside the tags from the conversation.
Desired Solution you'd like
Add a toggle to choose whether the reasoning should be kept or excluded from the conversation.
Alternatives Considered
No response
Additional Context
No response
@gaby commented on GitHub (Jun 14, 2025):
@Tureti Is this why the model goes into an infinite loop? I can't get it to work with OWUI.
I keeps "thinking" about the same thing in an endless loop.
@Tureti commented on GitHub (Jun 14, 2025):
@gaby I don't think that this has anything to do with the looping. If you are using a quantized version of the small model maybe this helps you. It's from the Qwen team about their own reasoning models, but might help:
@gaby commented on GitHub (Jun 14, 2025):
@Tureti I'm not using the GGUF version. I'm serving the model using the official instructions from
MistralwithvLLM.When prompting it from
Open-WebUIit goes into an infinite loop thinking.EDIT: It might be a
vLLMissue, this person reported the same issue https://huggingface.co/mistralai/Magistral-Small-2506/discussions/15@Tureti commented on GitHub (Jun 14, 2025):
@gaby Okay. As long as it does this in the first message as well (first one is unaffected by this open-webui behavior), it has nothing to do with this issue. If it only happens after multiple turns then it might be affected, but I'm honestly not sure if it could cause looping.