Diffusion Draft Models (DFlash)

[sammcj]

A rather neat approach of using a diffusion model as a draft to dramatically speed up inference is doing the rounds.

Blog: https://z-lab.ai/projects/dflash/
Paper: https://arxiv.org/html/2602.06036v1

@bstnxbt made a MLX server for it: https://github.com/bstnxbt/dflash-mlx

Qwen 3.5 DFlash (draft) models:

• https://huggingface.co/z-lab/Qwen3.5-4B-DFlash
• https://huggingface.co/z-lab/Qwen3.5-9B-DFlash
• https://huggingface.co/z-lab/Qwen3.5-35B-A3B-DFlash
• https://huggingface.co/z-lab/Qwen3.5-27B-DFlash

Testing out Qwen 3.5 27B 4-Bit, with DFlash draft on M5 Max: 1904 tokens | 40.0 tok/s | 82.6% acceptance

Looks significantly more promising than #500

Wondering how viable implementing diffusion draft models in oMLX might be?

[sammcj]

And I just saw 28fab9f 😂 well done!