Tools
MTPLX gets 3x MLX speedups on Qwen 3.8 27B using the model's native multi-token prediction head, no draft model needed
youssofal/MTPLX implements native MTP speculative decoding on Apple Silicon, reporting roughly 3x faster generation on Qwen 3.8 27B without any external drafter model. Eliminating the separate draft model removes the usual memory overhead and the draft-target alignment tuning that makes speculative decoding awkward on memory-constrained Macs. At 1,585 stars since 2026-05-02 and pushed 2026-08-22, it is smaller than the mainline MLX projects but targets a specific bottleneck they have not closed.
↳ Follow the thread