Adding Support for Qwen3.5 & Qwen3.5 MoE (Vision)

[johnmai-dev]

Proposed changes

1. Add Qwen3.5 and Qwen3.5 MoE (Vision)
2. Skip sanitize when metadata.format == "mlx" to avoid reapplying HF-only transforms (e.g., norm + 1) that cause gibberish output. https://github.com/Blaizzy/mlx-vlm/blob/a673abf2e91c0afbb0fd53d8cf6cd7eda9bd487f/mlx_vlm/utils.py#L215

Ported

• [MODEL] support qwen3.5 series w/o vision mlx-lm#869
• [MODEL] support qwen3.5 series Blaizzy/mlx-vlm#722

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[johnmai-dev]

mlx-community/Qwen3.5-27B-4bit

mlx-community/Qwen3.5-35B-A3B-4bit

[tpae]

Would love this to be merged soon @davidkoski 🙏

[davidkoski]

The new models look good but see my question on the weight loading and let me know what you think.

[johnmai-dev]

Qwen/Qwen3.5-0.8B

[johnmai-dev]

The new models look good but see my question on the weight loading and let me know what you think.

Previously, I referred to the implementation at Blaizzy/mlx-vlm -> utils.py/#L215, and at first I thought it was necessary to keep the logic consistent with mlx-vlm.

After reading your suggestion, I’ve already used loadArraysAndMetadata.

[johnmai-dev]

performance optimization

Before

After

[ronaldmannak]

Am I the only one checking the status of this PR every few hours? Can't wait @johnmai-dev :)

[johnmai-dev]

Am I the only one checking the status of this PR every few hours? Can't wait @johnmai-dev :)

haha,I can't wait too!This PR is ready.❤️

[davidkoski]

Looks great, thank you!