minimax-m3.1POST
Inco AI open-sources DFlash 2 speculative decoding for MiniMax M3.1 Flash
A free, commercially usable draft model for M3.1 Flash that keeps output identical while claiming up to 3.5x speedup; measured 3.53x on a single request with two GB300s, dropping to 1.6–2.6x at 8–32 concurrent requests.
墨无归
@MoblacksoilArchived 4 days agoOriginally posted 2026.10.23
墨无归 @MoblacksoilOriginally posted 2026.10.23
想让 MiniMax M3.1 Flash 跑得更快，Inco AI 开源了一个配套小模型 DFlash 2，免费开源、可商用，输出和原模型一致，不掉质量。 做法是让小模型先一口气猜出一段后续内容，再交给大模型核对，省掉一个字一个字往外蹦的时间。官方说最高快 3.5 倍；模型页给的条件是两张 GB300、一次只跑一个请求，实测快 3.53 倍；同时跑 8 到 32 个请求时，提速回落到约 1.6–2.6 倍。
This post quotes another post: “inco_ai open-sources DFlash 2 drafter for MiniMax M3.1 Flash under Apache 2.0”
x.com/moblacksoil/status/2113531367703134675Archived 2026.10.24 12:10