Qwen3.6-35B-A3B-MTP-4bit

This repository contains quantized Multi-Token Prediction (MTP) drafter weights split from Qwen/Qwen3.6-35B-A3B for use with mlx-vlm speculative decoding.

This is not a standalone chat or text-generation model. Load it as the draft model alongside a compatible Qwen3.6 35B-A3B target checkpoint.

Use with mlx-vlm

uv run mlx_vlm.generate \
  --model Qwen/Qwen3.6-35B-A3B \
  --draft-model mlx-community/Qwen3.6-35B-A3B-MTP-4bit \
  --prompt "Hi, how are you?" \
  --max-tokens 256 \
  --enable-thinking

For local weights:

uv run mlx_vlm.generate \
  --model /path/to/target-model \
  --draft-model /path/to/Qwen3.6-35B-A3B-mtp-4bit \
  --prompt "Hi, how are you?" \
  --max-tokens 256 \
  --enable-thinking

Model Details

Model type: qwen3_5_mtp
MTP block size: 2
Target architecture: Qwen3.6 35B-A3B
Precision: MLX affine 4-bit, group size 64
Runtime: MLX / mlx-vlm
Format: Safetensors with MLX-compatible config and tokenizer files

The stored tensors use MLX affine quantization as described in config.json.

Intended Use

Use this repo only as a speculative decoding drafter for compatible Qwen3.6 35B-A3B checkpoints. The target model verifies drafted tokens, while this MTP model proposes candidate tokens per decoding step.

Limitations

This checkpoint requires runtime support for Qwen MTP draft models in mlx-vlm. Standard standalone generation through generic Transformers APIs is not expected to work with this repository by itself.

Please refer to the upstream Qwen/Qwen3.6-35B-A3B model card and license terms for model usage constraints.

Loading…