DFlash makes Qwen3.6 27B 2.2x faster with no quality loss

DFlash accelerates Qwen3.6-27B inference by 2.2x on a single RTX 6000, from 44 to 98 tok/s, with no reported quality loss. The technique has a 30% token acceptance rate, compared to MTP's 71% at 1.45x speedup.