MiniMax M3.1 Flash hits >1K tok/s decode with day-0 SGLang support across NVIDIA and AMD
MiniMax_AI · x · 2026-10-23
MiniMax confirms M3.1 Flash ships with day-0 SGLang support, exceeding 1K tok/s at bs1 decode and production-ready from launch.
Key points:
- Broad hardware support: NVIDIA Blackwell/Hopper (GB300/B300/B200/H200) plus AMD Instinct MI350X/MI355X
- 100K total tok/s per GPU on the AgentX coding-agent trace with 96 concurrent sessions
- Speculative decoding via MiniMax's DSpark draft and a DFlash2 draft trained by radixark; HiCache and multimodal input supported
- MegaKernels deliver 2.15× faster single-request decode; joint effort by MiniMax, NVIDIA, AMD and radixark, with RL support via Miles
More from Infra
- vLLM ships Day-0 serving for MiniMax M3.1 Flash: NVFP4 cuts KV memory 43.75%, 2.7× concurrency — vllm_project · 2026-10-23
- vLLM details M3.1 Flash serving: NVFP4 cuts weights 44.6%, boosts capacity 2.7x — vllm_project · 2026-10-23
- M3.1 Flash hits 191.5k tok/s per GPU on Vera Rubin NVL72, leads open models on agent benchmarks — vllm_project · 2026-10-23
- Agents that don't close connections on retry waste ~148 tokens per task, vLLM tests show — No-Peanut-6988 · 2026-10-23
- Ex-Googler with 35 Years in Software Pursues Local AI Coding and Music Video Generation — julianharris · 2026-10-23
- Intelligence too cheap to meter: Opus price down 4x while AA score jumps 21 to 58 — haider1 · 2026-10-23