DFlash speeds up Qwen3.6 27B by 2.2x with no quality loss

Reddit benchmarks show DFlash achieving 98 tok/s on Qwen3.6-27B with a single RTX 6000, a 2.2x speedup over baseline. MTP (speculative decoding) reached 65 tok/s (1.45x) with a 71% acceptance rate.