[GH-ISSUE #12459] Qwen3-235B outputs <|endoftext|> tokens not specified in modelfile #34038

Closed
opened 2026-04-22 17:15:56 -05:00 by GiteaMirror · 2 comments
Owner

Originally created by @asitwere on GitHub (Sep 30, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/12459

Originally assigned to: @pdevine on GitHub.

What is the issue?

The Qwen3-235B model is occasionally outputting an <|endoftext|> token in some of its generated responses (particularly during lengthy outputs) and subsequently writing in circles.

This <|endoftext|> token does not appear in the standard modelfile config, which only includes <|im_start|> and <|im_end|>

Specific models being used:

  • qwen3:235b-a22b-instruct-2507-q8_0
  • qwen3:235b-a22b-thinking-2507-q8_0

Relevant log output

Relevant portion of example output showing the issue:

[...] At dawn, he returned to the dock, only to discover once more that no check or letter had come. He watched the mail boat depart, the islands reappearing through the mist.<|endoftext|>i need to summarize the provided text as requested

We are given a text to summarize [...]

OS

macOS

GPU

Apple

CPU

Apple

Ollama version

0.12.3

Originally created by @asitwere on GitHub (Sep 30, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/12459 Originally assigned to: @pdevine on GitHub. ### What is the issue? The Qwen3-235B model is occasionally outputting an `<|endoftext|>` token in some of its generated responses (particularly during lengthy outputs) and subsequently writing in circles. This `<|endoftext|>` token does not appear in the standard modelfile config, which only includes `<|im_start|>` and `<|im_end|>` Specific models being used: - qwen3:235b-a22b-instruct-2507-q8_0 - qwen3:235b-a22b-thinking-2507-q8_0 ### Relevant log output Relevant portion of example output showing the issue: [...] At dawn, he returned to the dock, only to discover once more that no check or letter had come. He watched the mail boat depart, the islands reappearing through the mist.<|endoftext|>i need to summarize the provided text as requested We are given a text to summarize [...] ### OS macOS ### GPU Apple ### CPU Apple ### Ollama version 0.12.3
GiteaMirror added the bug label 2026-04-22 17:15:56 -05:00
Author
Owner

@rick-github commented on GitHub (Sep 30, 2025):

#12444

<!-- gh-comment-id:3353710065 --> @rick-github commented on GitHub (Sep 30, 2025): #12444
Author
Owner

@pdevine commented on GitHub (Oct 1, 2025):

This is fixed now. You can pull the model again to grab the new template:

ollama pull qwen3:235b-a22b-thinking-2507-q8_0

This will only download the template and not the weights so should be fast.

<!-- gh-comment-id:3358512915 --> @pdevine commented on GitHub (Oct 1, 2025): This is fixed now. You can pull the model again to grab the new template: ``` ollama pull qwen3:235b-a22b-thinking-2507-q8_0 ``` This will only download the template and not the weights so should be fast.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#34038