Qwen3.6-35B-A3B: MLX 4-bit with native MTP

A 4-bit MLX checkpoint of Qwen/Qwen3.6-35B-A3B that keeps the model's native multi-token prediction (MTP) layer, so an engine that drafts with it gets everything from one repo.

What's in it

• The four model-*.safetensors shards and every config and tokenizer file are byte-identical to mlx-community/Qwen3.6-35B-A3B-4bit at revision 38740b8: affine 4-bit, group size 64, routers at 8 bits.
• mtp-4bit.safetensors adds the MTP layer that MLX conversions drop. It holds the official mtp.* weights from Qwen/Qwen3.6-35B-A3B at revision 995ad96, quantized the same way: affine 4-bit, group size 64, the router and shared-expert gate at 8 bits, norms in BF16. The tensors are named language_model.mtp.*.
• model.safetensors.index.json doesn't list the MTP file, so loaders that don't draft ignore it and load exactly the mlx-community model.

Use

With mlx-vlm, the same as the mlx-community conversion:

pip install -U mlx-vlm
python -m mlx_vlm.generate --model Vontra/Qwen3.6-35B-A3B-MLX-4bit-MTP --max-tokens 100 --temperature 0.0 --prompt "Describe this image." --image <path_to_image>

TensorFold drafts with the MTP layer and verifies every drafted token against the model, so its output equals the model's own decoding. Support for this model there is in development.

License

Apache-2.0, from Qwen/Qwen3.6-35B-A3B. The model is by the Qwen team; the MLX conversion of the main weights is by mlx-community.