Key Info

MiniMax M3.1 Flash reaches over 1K tok/s at batch size 1 decode with SGLang, with production-ready day-0 support from SGLang and Miles.

Highlights

  • Over 1K tok/s at bs1 decode using SGLang
  • Broad hardware support including NVIDIA Blackwell and Hopper (GB300/B300/B200) and AMD
  • Joint effort by MiniMax, NVIDIA, AMD and RadixArk
  • Reinforcement learning supported via Miles
  • Covers both inference and RL from day one