Deepseek 4.1, Qwen and GLM 5.3 flash don't load!

[Metal-Gear-Solid]

Hello @Blaizzy

My man, nothing loads, what's going on? I came in late on your release, congrats, but I thought everything would be working by now?

Let me list the models that don't work, no matter which version of it I tried:

• DeepSeek V4-exp and vision variant
• DeepSeek V4.1, literally tried three separate variants
• Qwen 3.8 Flash Next, tried 5 versions
• GLM 5.3 Flash (this one was the oQ4e-MTP I had around, did nto have any other variant for it on storage)

The only model out of my 20+ that work is literally:

• DeepSeek V4-0731-MXFP4-MLX
• Qwen 3.5 out of retirement to see if anything runs

It is a sweet app but at least let us know what does not work in app maybe?

Any plans to support Deepseek v4 vision and v4.1? I am super keen on this one "DeepSeek-V4.1-Flash-Uncensored-MLX"

🦖 Version 0.3.8 (202609141816)

[linear-code]

NEY-196

[Lazarus-931]

hi @Metal-Gear-Solid thanks for detailed issue.

we're sorry you're having issues and planning on addressing these.

As for the models,

1. v4-exp -> was merged for and should be in nativ 0.3.8: Add DeepSeek-V4 Flash Vision EXP support mlx-vlm#2136, can you file a new issue with diagnostics or what you see in chat so we can take a look?
2. v4.1 -> pr hasn't been merged yet(DeepSeek V4.1 vision and language model support mlx-vlm#2214), will be delivered in next nativ release
3. As for flash next, support was merged into mlx-vlm and should work in nativ: Add Qwen3.8-Flash-Next support mlx-vlm#2032, please file a new issue with diagnostics if you get checkpoint errors
4. For the 5.3 flash, we currently haven't added support, spec decode is being re-designed (Rebuild speculative decoding around request caches and shared prefill mlx-vlm#2206), so not only that but mtp for dsv4.1 as well hasn't been added, will be soon.

[Lazarus-931]

following up @Metal-Gear-Solid, I tried vision-exp and qwen3.8-next checkpoints and they do work on mlx-vlm, so curious what issue you're running into

here are the checkpoints on hf as well:

1. https://huggingface.co/mlx-community/DeepSeek-V4-Flash-Vision-Exp-MLX
2. https://huggingface.co/mlx-community/Qwen3.8-Flash-Next-4bit

[stardomains3]

Anyone know if mlx-community/Qwen-Image-2.1-MLX-4bit works with it?

Also it would be cool to open the Discussions tab in Github for Nativ so folks could figure stuff out together.

[Lazarus-931]

Anyone know if mlx-community/Qwen-Image-2.1-MLX-4bit works with it?

Also it would be cool to open the Discussions tab in Github for Nativ so folks could figure stuff out together.

hi @stardomains3, we have a finished pr in mlx-vlm to be merged for the upcoming weekly release: Blaizzy/mlx-vlm#2321

as for discussion tab, will look into it :)

[Lazarus-931]

discussions page is open as well!

[Blaizzy]

Hey @Metal-Gear-Solid

Thanks for reaching out!

Indeed, we are working on ways to make it clear on the app what's not yet supported.

To clarify, we deliberately delayed those models (DeepSeek, Qwen3.8) because we are working on the best experience possible.

It will be worth the wait!