Space Bunny（M3.1 Flash 匿名版）揭晓：思考默认 max 且无法关闭

Cuth @ItsCuthulhu原推发布于 2026.10.22

在 X 上打开原推

Space Bunny has a name now: it is MiniMax-M3.1-Flash, and it is live on MiniMax's API. The stealth model people were testing blind in September turned out to be MiniMax's new flagship flash model. It is multimodal, built for coding, and has a 1 million token context. The numbers: $0.15 per 1M input tokens and $0.60 output, with 74% on DeepSWE. Open weights land Oct 23. The interesting knob is thinking depth. You can set effort from low to max, but thinking is always on and defaults to max, so you cannot fully turn it off. The catch: it is API only until the weights drop, and always-on thinking at max default could get expensive on long sessions before you notice. Worth trying for long coding sessions once the weights land, or does always-on thinking at max default worry you about cost? https://x.com/MiniMax_AI/status/2112111925975032241 https://platform.minimax.io/docs/guides/text-generation

这条是引用着另一条写的：「MiniMax 官方预告 M3.1 Flash 发布，点名 Space Bunny」