DFlash speeds up local Qwen3.6-27B model 2.2x with no quality loss
A test on a single RTX 6000 GPU compared three ways of running the Qwen3.6-27B: baseline, MTP, and DFlash. The tasks were writing quicksort code, generating a Steam library in JSON format, solving a logic puzzle, and writing a sci-fi story. Baseline ran at 44 (1x). MTP ran at 65 (1.45x) with a 71% guess-.
DFlash ran at 98 (2.2x) with a lower 30% . DFlash predicts 15 tokens in a row at once, so it excels on repetitive, like JSON, where long correct guess-streaks are likely (152 , 3.4x there). On creative writing, most guesses are wrong, so the wasted prediction work can actually push speed below baseline (42 vs 44 ). MTP only guesses 3 tokens in parallel from within the model itself, so a wrong guess costs almost nothing, meaning it never drops below baseline speed.
All three methods produce identical final output. The conclusion: DFlash suits coding-style tasks best, while MTP suits chat or creative writing better.
Key points
- Compared baseline, MTP, and DFlash decoding on Qwen3.6-27B using one RTX 6000 GPU
- DFlash is fastest at 98 tokens/sec (2.2x) but has a low 30% guess-
- MTP runs at 65 tokens/sec (1.45x) with a steadier 71% and never falls below baseline
- Structured, repetitive text (like JSON) favors DFlash; creative writing favors MTP
- All three methods produce the exact same final output quality