github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 08:30:05 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #12234] qwen3:4b still output thinking progress with Chat API "think = false". #54652

New Issue
Closed
opened 2026-04-29 06:45:53 -05:00 by GiteaMirror · 10 comments
GiteaMirror commented 2026-04-29 06:45:53 -05:00
Owner
Copy Link

Originally created by @owenzhao on GitHub (Sep 9, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/12234

What is the issue?

qwen3:latest, which is 8b works well. However, qwen3:4b still output think progress in content.

Image

Relevant log output

OS

macOS

GPU

Apple

CPU

Apple

Ollama version

0.11.10

Originally created by @owenzhao on GitHub (Sep 9, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/12234 ### What is the issue? qwen3:latest, which is 8b works well. However, qwen3:4b still output think progress in content. <img width="770" height="380" alt="Image" src="https://github.com/user-attachments/assets/8bf16da8-55a6-452b-aee4-3f737c621533" /> ### Relevant log output ```shell ``` ### OS macOS ### GPU Apple ### CPU Apple ### Ollama version 0.11.10
GiteaMirror added the bug label 2026-04-29 06:45:53 -05:00
GiteaMirror commented 2026-04-29 06:45:56 -05:00
Author
Owner
Copy Link

@owenzhao commented on GitHub (Sep 9, 2025):

I think this probably is a regression of Ollama as in my memory qwen3:4b was good when think parameter was first introduced.

<!-- gh-comment-id:3272339768 --> @owenzhao commented on GitHub (Sep 9, 2025): I think this probably is a regression of Ollama as in my memory qwen3:4b was good when think parameter was first introduced.
GiteaMirror commented 2026-04-29 06:46:04 -05:00
Author
Owner
Copy Link

@owenzhao commented on GitHub (Sep 9, 2025):

Also, "stream = false". If that is relevant.

<!-- gh-comment-id:3272345787 --> @owenzhao commented on GitHub (Sep 9, 2025): Also, "stream = false". If that is relevant.
GiteaMirror commented 2026-04-29 06:46:05 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 9, 2025):

#12022

<!-- gh-comment-id:3272350639 --> @rick-github commented on GitHub (Sep 9, 2025): #12022
GiteaMirror commented 2026-04-29 06:46:08 -05:00
Author
Owner
Copy Link

@owenzhao commented on GitHub (Sep 9, 2025):

#12022

Thank you. However, it seems there is only one qwen3:4b on ollama.com. Can I get qwen3:4b-instruct in ollama.com. Or I have to get it from other place?

<!-- gh-comment-id:3272398350 --> @owenzhao commented on GitHub (Sep 9, 2025): > [#12022](https://github.com/ollama/ollama/issues/12022) Thank you. However, it seems there is only one qwen3:4b on ollama.com. Can I get qwen3:4b-instruct in ollama.com. Or I have to get it from other place?
GiteaMirror commented 2026-04-29 06:46:10 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 9, 2025):

https://ollama.com/library/qwen3/tags

<!-- gh-comment-id:3272401635 --> @rick-github commented on GitHub (Sep 9, 2025): https://ollama.com/library/qwen3/tags
GiteaMirror commented 2026-04-29 06:46:12 -05:00
Author
Owner
Copy Link

@owenzhao commented on GitHub (Sep 9, 2025):

I also tested with 0.6b and 1.7b, they both work well most of the time. However, some times 1.7b will append "no_think" to answers. For example:

When translating "PROXY: %lld", you get "代理：%lld /no_think".

<!-- gh-comment-id:3272413460 --> @owenzhao commented on GitHub (Sep 9, 2025): I also tested with 0.6b and 1.7b, they both work well most of the time. However, some times 1.7b will append "no_think" to answers. For example: When translating "PROXY: %lld", you get "代理：%lld /no_think".
GiteaMirror commented 2026-04-29 06:46:18 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 9, 2025):

When translating "PROXY: %lld", you get "代理：%lld /no_think".

This a side effect of ollama trying to control the thinking mode. The ollama template controls thinking by adding /think or /no_think to the end of the last user message. If that message is something that the model is going to send back to the client (eg when translating something), then it's possible that the control token will be sent as well. You can work around this by moving the code that adds the control token to the part of the template that creates the system message. The only drawback is that the user message can then override the system message and toggle the thinking flag.

This may be why qwen decided to move away from a hybrid model for the update to qwen3:4b.

<!-- gh-comment-id:3272528397 --> @rick-github commented on GitHub (Sep 9, 2025): > When translating "PROXY: %lld", you get "代理：%lld /no_think". This a side effect of ollama trying to control the thinking mode. The [ollama template](https://ollama.com/library/qwen3:1.7b/blobs/ae370d884f) controls thinking by adding `/think` or `/no_think` to the end of the last user message. If that message is something that the model is going to send back to the client (eg when translating something), then it's possible that the control token will be sent as well. You can work around this by moving the code that adds the control token to the part of the template that creates the system message. The only drawback is that the user message can then override the system message and toggle the thinking flag. This may be why qwen decided to move away from a hybrid model for the update to qwen3:4b.
GiteaMirror commented 2026-04-29 06:46:19 -05:00
Author
Owner
Copy Link

@owenzhao commented on GitHub (Sep 10, 2025):

Thanks for your explanation. @rick-github

However, I am not using "no_think" to control. I have used the Ollama's API with "think = false". I knew that "no_think" in prompt worked before. But since Ollama added "think" parameter to Ollama own API, I have turned to this.

I intended use Ollama Native API instead of the compatible OpenAI API, as the latter doesn't support "think" parameter now.

<!-- gh-comment-id:3272938140 --> @owenzhao commented on GitHub (Sep 10, 2025): Thanks for your explanation. @rick-github However, I am not using "no_think" to control. I have used the Ollama's API with "think = false". I knew that "no_think" in prompt worked before. But since Ollama added "think" parameter to Ollama own API, I have turned to this. I intended use Ollama Native API instead of the compatible OpenAI API, as the latter doesn't support "think" parameter now.
GiteaMirror commented 2026-04-29 06:46:19 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 10, 2025):

That is how the think API works. The template adds /think or /no_think based on the value of the flag passed in the API call.

<!-- gh-comment-id:3273015623 --> @rick-github commented on GitHub (Sep 10, 2025): That is how the `think` API works. The template adds `/think` or `/no_think` based on the value of the flag passed in the API call.
GiteaMirror commented 2026-04-29 06:46:19 -05:00
Author
Owner
Copy Link

@owenzhao commented on GitHub (Sep 10, 2025):

That is how the think API works. The template adds /think or /no_think based on the value of the flag passed in the API call.

Really? In that case, only qwen3-4b-instruct is recommended as only it follows no_thinking very well but in fact it doesn't support thinking at all. It humors me somehow.

<!-- gh-comment-id:3273347566 --> @owenzhao commented on GitHub (Sep 10, 2025): > That is how the `think` API works. The template adds `/think` or `/no_think` based on the value of the flag passed in the API call. Really? In that case, only qwen3-4b-instruct is recommended as only it follows no_thinking very well but in fact it doesn't support thinking at all. It humors me somehow.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label bug
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#54652
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?