mirror of
https://github.com/open-webui/open-webui.git
synced 2026-05-07 03:18:23 -05:00
[GH-ISSUE #21348] issue: v0.8.0 reasoning trace is visually split to many parts, causing the browser to slow down to a halt #58116
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: github-starred/open-webui#58116
Reference in New Issue
Block a user
Delete Branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @rotemdan on GitHub (Feb 13, 2026).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/21348
Check Existing Issues
Installation Method
Docker
Open WebUI Version
v0.8.0
Ollama Version (if applicable)
No response
Operating System
Windows 11
Browser (if applicable)
Brave (latest)
Confirmation
README.md.Expected Behavior
Regular reasoning trace rendering.
Actual Behavior
Seems to be some sort of rendering issue with reasoning trace possibly split at each token output. I think the screen capture is the best way to show this.
Steps to Reproduce
Prompt a model that outputs a reasoning trace.
Logs & Screenshots
Additional Information
Never seen the issue before v0.8.0.
@rotemdan commented on GitHub (Feb 13, 2026):
To clarify I'm seeing this with my own local backend (called NodeLM).
It happens only on reasoning traces, not on non-reasoning responses.
I tried to see if I get this with OpenRouter and it doesn't seem so (at the moment, based on limited testing).
Something about the stream response my backend is generating is causing
v0.8.0to produce this behavior, while previous versions of open-webui didn't. It's possible other backends also cause that, but I'm not able to test all of them at the moment.Edit: I believe my backend produces standard streaming outputs. Speculation: it may actually be related to connection latency or the fact the server is local.
I'll need to investigate what exactly is causing it.
@LostVector commented on GitHub (Feb 13, 2026):
getting this too ... so it wasn't always like this? lol
@Yundin commented on GitHub (Feb 13, 2026):
Having the same issue after the 0.8.0 update
I'm using a pipe function to OpenRouter. I've checked that its output hasn't changed: it yields
<think>\nblock and reasoning tokens after that, followed by\n</think>\n\nand the rest of the content. It is visual representation that changed.@spectre-pro commented on GitHub (Feb 13, 2026):
I use gemini 3 pro, also have this problem
@sailstudio commented on GitHub (Feb 13, 2026):
I use minimax-m2.1(local host vllm provider), also have this problem, and i can not revert to last verison (database changed)
@ByzantineProcess commented on GitHub (Feb 13, 2026):
Also getting this issue with an Ollama backend using deepseek-r1 and qwen3. gpt-oss and glm-4.7-flash don't seem to be affected?
@RedBlizard commented on GitHub (Feb 13, 2026):
I have the same problem for now rolling back to release v0.7.2 fixed my problem ! and removed watchtower in my docker-compose for now ! i hope this will be fixed soon.
@spectre-pro commented on GitHub (Feb 13, 2026):
database be changed also can rolling back?
@kleymenus commented on GitHub (Feb 13, 2026):
I have this issue, too... w. OpenAI API (Mistral, Google), Ollama
@Classic298 commented on GitHub (Feb 13, 2026):
During extensive testing we DID see this behavior but only for the minimax api.
We never saw this behavior on OpenAI API, nor OpenRouter, nor LiteLLM, nor OpenAI, nor Vertex AI
I assumed minimax api due to the very new model release had issues and i couldn't reproduce it on openrouter at all.
What api endpoints are you guys using?
There has to be a difference in how these Providers stream their responses.
@kleymenus commented on GitHub (Feb 13, 2026):
@Classic298 for instance:
https://api.mistral.ai/v1
https://generativelanguage.googleapis.com/v1beta/openai
and Ollama
(in previous versions, everything worked with the same initial settings)
@RedBlizard commented on GitHub (Feb 13, 2026):
I don't use external api's just Ollama with http://host.docker.internal:11434
@Classic298 commented on GitHub (Feb 13, 2026):
Ok will try with generative language openai endpoint by Google. Give me a minute.
@kleymenus commented on GitHub (Feb 13, 2026):
@Classic298 just a note: reasoning tags could be inside the stream of the main assistant's response
@Omaha2002 commented on GitHub (Feb 13, 2026):
Same here, not on Mistral API, OpenAI API , Grok API or Anthropic API, they all act "normal".
Qwen3 running on vLLM also shows every second "Thought for less than one second...".
In advanced parameters I added custom Reasoning tags for Qwen that worked up until 0.8.0.
For now i disabled them and see the reasoning.
@Classic298 commented on GitHub (Feb 13, 2026):
I cannot reproduce via generativelanguage.googleapis.com/v1beta/openai endpoint - but maybe because they don't send the reasoning traces at all.
@Classic298 commented on GitHub (Feb 13, 2026):
and other endpoints i can try that dont work? unfortunately i dont have mistral
@kleymenus commented on GitHub (Feb 13, 2026):
@Omaha2002 @Classic298
Google hiding reasoning tags by default on this API type.
for the experiment, add this to your prompt at the very end for the Mistral API or Google API:
@Classic298 commented on GitHub (Feb 13, 2026):
Are you sure a prompt is going to override the APIs behaviour of not sending internal reasoning?
@spectre-pro commented on GitHub (Feb 13, 2026):
Google hiding reasoning tags by default on this API type.
you need to add Advanced Params
@Classic298 commented on GitHub (Feb 13, 2026):
thanks, i will try that
@Classic298 commented on GitHub (Feb 13, 2026):
thanks reproducible with that. Will investigate.
@kleymenus commented on GitHub (Feb 13, 2026):
@Classic298 this is to emulate the general response and problem - the model begins to incorporate thought within the main block itself, yes. so, even if you don't have Mistral or any other api access - you will reproduce an issue in generic way
@spectre-pro commented on GitHub (Feb 13, 2026):
So can I roll back? I'm worried that doing so will corrupt my database.
@Classic298 commented on GitHub (Feb 13, 2026):
You can roll back if you also roll back your database from a backup
The migrations have moved your access control, chat messages and prompts into new tables
@spectre-pro commented on GitHub (Feb 13, 2026):
Unfortunately, I didn't back up the database before the update.
@Classic298 commented on GitHub (Feb 13, 2026):
If this is a backend issue (might as well be) then everyone here can easily apply the changes without needing to wait for a new release or even for the PR to get merged. All installation methods allow you to change the backend python files, restart, and done.
PS: Just generally said - we encourage everyone to have a separate dev environment. Either for actually using it on a day-2-day basis (for single user setups) or as a testing environment and to test upcoming releases for your production use. If more people can help test the
devbranch, things like this could be caught earlier. We do a lot of testing already, but as evident, despite testing many different providers - we can't try them all.@Classic298 commented on GitHub (Feb 13, 2026):
Might have a fix ready. Give me a little bit to verify.
@spectre-pro commented on GitHub (Feb 13, 2026):
ok thx
@Classic298 commented on GitHub (Feb 13, 2026):
Hell yeah
Fix works with generativelanguages endpoint and minimax endpoint and doesnt break openrouter (which previously worked)
so looks goooood
@spectre-pro commented on GitHub (Feb 13, 2026):
Where or when can we get this fix works
@Classic298 commented on GitHub (Feb 13, 2026):
@spectre-pro give me a minute 👀 patience
@spectre-pro commented on GitHub (Feb 13, 2026):
ok...
@Classic298 commented on GitHub (Feb 13, 2026):
anyone with this issue i encourage you to manually apply these changes to your open webui (before anyone asks, yes you can also modify the docker image, just docker exec and bash into it to modify it) then restart after modifying your backend and try it out
https://github.com/open-webui/open-webui/pull/21355
@Classic298 commented on GitHub (Feb 13, 2026):
Please come back and confirm working (or not) when done.
Additionally, i ran 8 different tests on this change and they all passed - so in addition of my own testing, i am highly confident this should work for everyone
@kleymenus commented on GitHub (Feb 13, 2026):
fixed for me 👍
@spectre-pro commented on GitHub (Feb 13, 2026):
it works for me
@Classic298 commented on GitHub (Feb 13, 2026):
perfect thanks for confirming guys.
Reminder: Just generally said - we encourage everyone to have a separate dev environment. Either for actually using it on a day-to-day basis (for single user setups) or as a testing environment and to test upcoming releases for your production use. If more people can help test the dev branch, things like this could be caught earlier. We do a lot of testing already, but as evident, despite testing many different providers - we can't try them all.
Will leave this open until the PR is merged.
@RedBlizard commented on GitHub (Feb 13, 2026):
@Classic298 thanks for your support , but i have to wait until it is merged with next release.
@Classic298 commented on GitHub (Feb 13, 2026):
@RedBlizard technically you don't have to. You can just apply these changes manually to the backend files.
@sailstudio commented on GitHub (Feb 13, 2026):
@Classic298 I use this patch，and fix think tag issue , nice work!
but display maybe incorrect when the "think" process contains a <code_interpreter> block, please check this:
@Classic298 commented on GitHub (Feb 13, 2026):
for the code interpreter, why should it render it as html? it is always json
what model is this? I dont get the non-detected think tags anymore. Also check your advanced model settings, did you define any custom think tags there?
@sailstudio commented on GitHub (Feb 13, 2026):
@Classic298 commented on GitHub (Feb 13, 2026):
@sailstudio could you provide the raw response text for me please - just using the copy button below
and please also say if you configured custom think tags or not.
@Classic298 commented on GitHub (Feb 13, 2026):
nevermind could reproduce it and i have a fix for it . 1min
@spectre-pro commented on GitHub (Feb 13, 2026):
I think the reason is that the model invokes the tool while thinking, rather than after thinking ends.
This caused the tool to break the thinking area.
@Classic298 commented on GitHub (Feb 13, 2026):
thats not the issue.
@Classic298 commented on GitHub (Feb 13, 2026):
@sailstudio pushed a new fix to the PR
https://github.com/open-webui/open-webui/pull/21355
apply all changes again
@tjbck commented on GitHub (Feb 13, 2026):
Could anyone confirm if this issue has been resolved in dev?
@Classic298 commented on GitHub (Feb 13, 2026):
yes resolved in dev