ollama - 💡(How to fix) Fix [0.21.0-rc0] qwen3.6:35b-a3b-nvfp4 fails to load: layer 0 missing linear attention projections [1 participants]

Error: 500 Internal Server Error: mlx runner failed: Error: layer 0: missing linear attention projections (exit: exit status 1)

TL;DR

The issue is likely due to a compatibility problem between the qwen3.6:35b-a3b-nvfp4 model and the current Ollama version, causing a failure to load the model due to missing linear attention projections.

Guidance

• Verify that the qwen3.6:35b-a3b-nvfp4 model is compatible with Ollama version 0.21.0-rc0 by checking the model's documentation or release notes.
• Test other variants of the model, such as mxfp8 or mlx-bf16, to see if the issue is specific to the nvfp4 variant.
• Check for any updates or patches for Ollama that may address compatibility issues with the qwen3.6:35b-a3b-nvfp4 model.
• Consider providing additional logs or output to help diagnose the issue, as the current error message is quite general.

Notes

The issue may be specific to the Apple Silicon hardware or macOS version, so testing on a different environment may help isolate the problem.

Recommendation

Apply workaround: Test other variants of the model, such as mxfp8 or mlx-bf16, to see if the issue is specific to the nvfp4 variant. This may help identify if the problem is with the model or the Ollama version.