Qwen3.6-35B-A3B-DFlash2-4bit

The DFlash 2 draft model for Qwen3.6-35B-A3B, as a standard safetensors checkpoint with MLX affine 4-bit weights (group size 64).

Source and changes

• Source: the draft/ files of incoai/Qwen3.6-35B-A3B-Splash, revision 0f4714b2db37b5f3c42a10de07281e74f88e4adc (Apache-2.0). That package names its draft source as incoai/Qwen3.6-35B-A3B-DFlash2, revision 8e713508f0bb02f03b5cb5cabbc8d9604f924be2.
• The draft model was made by Inco AI. This repository is not published or endorsed by Inco AI.
• Changes: the Splash runtime storage (tiled q4 integers with bf16 scales and biases) was converted to row-major MLX affine storage in one model.safetensors, and a config.json was written. The integers, scales and biases are moved bit for bit; no tensor was requantized (208 tensors). The original unquantized weights are not recovered. Mask token id and RoPE parameters come from the target's configuration. source.json lists the SHA-256 of every source file.

Use with vLLM on Apple silicon (exo-vllm plugin)

vllm serve mlx-community/Qwen3.6-35B-A3B-4bit \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml --reasoning-parser qwen3 --kv-cache-dtype fp8 \
  --speculative-config '{"method": "dflash", "model": "mlx-community/Qwen3.6-35B-A3B-DFlash2-4bit", "num_speculative_tokens": 7}'

License

Apache-2.0, as the source package. See LICENSE.