Paper Notes: dFlash
last updated 2026-07-13
dFlash is a block-causal diffusion model for speculative decoding. Unlike EAGLE/EAGLE-2/EAGLE-3, it generates its draft tokens in parallel, so you don’t have to pay an autoregressive cost.
This allows dFlash models to be even larger than EAGLE models, and thus possibly more accurate. However, they do suffer from diffusion limitations.
DeepSeek’s future dSpark gets around this by doing mostly parallel prediction, with a light autoregressive head on top.
It’s a popular replacement for EAGLE.