[GH-ISSUE #12569] In Ollama v0.12.3 and v0.12.5, the deepseek-r1:8b model outputs are chaotic, and its reasoning capabilities have significantly degraded, showing a serious performance regression compared to v0.11.11. #8337

Closed
opened 2026-04-12 20:55:20 -05:00 by GiteaMirror · 5 comments
Owner

Originally created by @minghua-123 on GitHub (Oct 11, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/12569

What is the issue?

Description:
After upgrading Ollama from v0.11.11 to v0.12.3 and v0.12.5, when using the deepseek-r1:8b model for inference, the model's output shows significant abnormalities, manifested as:

The output content is semantically chaotic and logically fragmented, failing to form coherent responses;
The error rate for simple reasoning tasks (such as mathematical calculations and general knowledge Q&A) has increased significantly;
Compared to the v0.11.11 version, the model's response quality has sharply declined, making it almost unusable for practical tasks.

Relevant log output


OS

Windows

GPU

Nvidia

CPU

Intel

Ollama version

0.11.11、0.12.3、0.12.5

Originally created by @minghua-123 on GitHub (Oct 11, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/12569 ### What is the issue? Description: After upgrading Ollama from v0.11.11 to v0.12.3 and v0.12.5, when using the deepseek-r1:8b model for inference, the model's output shows significant abnormalities, manifested as: The output content is semantically chaotic and logically fragmented, failing to form coherent responses; The error rate for simple reasoning tasks (such as mathematical calculations and general knowledge Q&A) has increased significantly; Compared to the v0.11.11 version, the model's response quality has sharply declined, making it almost unusable for practical tasks. ### Relevant log output ```shell ``` ### OS Windows ### GPU Nvidia ### CPU Intel ### Ollama version 0.11.11、0.12.3、0.12.5
GiteaMirror added the bug label 2026-04-12 20:55:21 -05:00
Author
Owner

@jmorganca commented on GitHub (Oct 11, 2025):

Hi there, this should be fixed in 0.12.5. Sorry about that, and let me know if you're still seeing it.

<!-- gh-comment-id:3392836982 --> @jmorganca commented on GitHub (Oct 11, 2025): Hi there, this should be fixed in 0.12.5. Sorry about that, and let me know if you're still seeing it.
Author
Owner

@minghua-123 commented on GitHub (Oct 11, 2025):

@jmorganca

In version 0.11.11, reasoning during model thinking works correctly, as shown in the picture below.
Image

However, in versions 0.12.3 and 0.12.5, the reasoning and thinking are chaotic, and the output formatting is messed up.
Image

<!-- gh-comment-id:3392838347 --> @minghua-123 commented on GitHub (Oct 11, 2025): @jmorganca In version 0.11.11, reasoning during model thinking works correctly, as shown in the picture below. <img width="1187" height="2960" alt="Image" src="https://github.com/user-attachments/assets/dee46118-7df8-469a-9c94-dd755125e4b1" /> However, in versions 0.12.3 and 0.12.5, the reasoning and thinking are chaotic, and the output formatting is messed up. <img width="1231" height="2538" alt="Image" src="https://github.com/user-attachments/assets/4167c206-ae21-4418-b3e4-2b748c91eeb6" />
Author
Owner

@minghua-123 commented on GitHub (Oct 11, 2025):

您好,这应该在 0.12.5 中修复。对此很抱歉,如果您仍然看到它,请告诉我。

在0.12.5中仍然出现此问题https://github.com/ollama/ollama/issues/12569#issuecomment-3392838347

Image
<!-- gh-comment-id:3392840915 --> @minghua-123 commented on GitHub (Oct 11, 2025): > 您好,这应该在 0.12.5 中修复。对此很抱歉,如果您仍然看到它,请告诉我。 在0.12.5中仍然出现此问题https://github.com/ollama/ollama/issues/12569#issuecomment-3392838347 <img width="1343" height="1282" alt="Image" src="https://github.com/user-attachments/assets/ba94366d-af04-44a7-b915-d963a1f72dbe" />
Author
Owner

@minghua-123 commented on GitHub (Oct 11, 2025):

您好,这应该在 0.12.5 中修复。对此很抱歉,如果您仍然看到它,请告诉我。

@jmorganca
0.12.5 exhibits more chaotic reasoning compared to 0.12.3, with a large amount of meaningless repetition and negative words, in both Chinese and English conversations.

<!-- gh-comment-id:3392843826 --> @minghua-123 commented on GitHub (Oct 11, 2025): > 您好,这应该在 0.12.5 中修复。对此很抱歉,如果您仍然看到它,请告诉我。 @jmorganca 0.12.5 exhibits more chaotic reasoning compared to 0.12.3, with a large amount of meaningless repetition and negative words, in both Chinese and English conversations.
Author
Owner

@pdevine commented on GitHub (Oct 13, 2025):

There ended up being an issue w/ RoPE/Yarn in the default deepseek-r1 distill which was causing bad output. We switched back to running the deepseek-r1 distill on the legacy engine in 0.12.5, but there is how a fix (#12603) which will turn this back on in the new engine in the next release (0.12.6).

<!-- gh-comment-id:3398989750 --> @pdevine commented on GitHub (Oct 13, 2025): There ended up being an issue w/ RoPE/Yarn in the default deepseek-r1 distill which was causing bad output. We switched back to running the deepseek-r1 distill on the legacy engine in `0.12.5`, but there is how a fix (#12603) which will turn this back on in the new engine in the next release (`0.12.6`).
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#8337