From Chen et al. (arXiv 2602.06036), DFlash trains a lightweight drafter to predict blocks of tokens—typically 16 at a time—while the full target model verifies the block in one forward pass and accepts a greedy prefix. Implementations such as dflash-mlx on Apple Silicon and Meta's Muse Glimmer drafter use it to raise decode throughput without changing greedy output quality when verification is lossless.