[GH-ISSUE #10741] Ollama 0.7.0 and gemma3:27b-it-qat fail to process images generating internal error #53568

Open
opened 2026-04-29 03:49:05 -05:00 by GiteaMirror · 2 comments
Owner

Originally created by @stannenb on GitHub (May 16, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/10741

What is the issue?

When processing an image with a simple prompt like "what's in this picture? /Users/xxx/Downloads/Photo_092807_007.jpg", ollama 0.7.0 along with a freshly downloaded gemma3:27b-it-qat consistently produces gibberish:

Added image '/Users/xxx/Downloads/Photo_092807_007.jpg'
hSFi_xFsI_ l

</i>

</i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i>

See error log below for the consistent error this produces.
In contrast, qwen2.5vl:latest processes it correctly.

Relevant log output

[GIN] 2025/05/16 - 11:35:51 | 200 |  4.355901125s |       127.0.0.1 | POST     "/api/generate"
ggml_metal_graph_compute: command buffer 1 failed with status 5
error: Internal Error (0000000e:Internal Error)

OS

macOS

GPU

Apple

CPU

Apple

Ollama version

0.7.0

Originally created by @stannenb on GitHub (May 16, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/10741 ### What is the issue? When processing an image with a simple prompt like "what's in this picture? /Users/xxx/Downloads/Photo_092807_007.jpg", ollama 0.7.0 along with a freshly downloaded gemma3:27b-it-qat consistently produces gibberish: ``` Added image '/Users/xxx/Downloads/Photo_092807_007.jpg' hSFi_xFsI_ l </i> </i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i></i> ``` See error log below for the consistent error this produces. In contrast, qwen2.5vl:latest processes it correctly. ### Relevant log output ```shell [GIN] 2025/05/16 - 11:35:51 | 200 | 4.355901125s | 127.0.0.1 | POST "/api/generate" ggml_metal_graph_compute: command buffer 1 failed with status 5 error: Internal Error (0000000e:Internal Error) ``` ### OS macOS ### GPU Apple ### CPU Apple ### Ollama version 0.7.0
GiteaMirror added the bugmacos labels 2026-04-29 03:49:06 -05:00
Author
Owner

@es1o commented on GitHub (May 21, 2025):

I have the same problem on Mac Studio with M2Max. It runs fine on Mac Pro with M2Pro.

<!-- gh-comment-id:2898060681 --> @es1o commented on GitHub (May 21, 2025): I have the same problem on Mac Studio with M2Max. It runs fine on Mac Pro with M2Pro.
Author
Owner

@stannenb commented on GitHub (May 24, 2025):

This problem continues with 0.7.1 albeit with different behavior.

You still get the
ggml_metal_graph_compute: command buffer 1 failed with status 5 error: Internal Error (0000000e:Internal Error)

error, but the model then provides an analysis of the picture that is quite wrong. Two example: a low quality image of some desert terrain is interpreted as abstract imagery, and a poster of a man is interpreted as the closeup of an eye.

I, too, am using a Mac Studio with an Apple M2 Max.

<!-- gh-comment-id:2906938568 --> @stannenb commented on GitHub (May 24, 2025): This problem continues with 0.7.1 albeit with different behavior. You still get the `ggml_metal_graph_compute: command buffer 1 failed with status 5 error: Internal Error (0000000e:Internal Error)` error, but the model then provides an analysis of the picture that is quite wrong. Two example: a low quality image of some desert terrain is interpreted as abstract imagery, and a poster of a man is interpreted as the closeup of an eye. I, too, am using a Mac Studio with an Apple M2 Max.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#53568