Benchmarks show DFlash in llama.cpp boosts Qwen 3.6 27B inference up to 6x

On a single RTX 6000 Pro, DFlash achieved up to 3.7x speedups on coding tasks with no quality loss, compared to 2.7x for MTP and 6x for n-gram stack. The technique works at long contexts up to 36K tokens, as tested with Qwen3.6-27B.