MiniMax open-weights 428B-parameter M3.1 Flash with 23B active per token
lmoroney · x · 2026-10-24
MiniMax has released open weights for M3.1 Flash — the model previously teased as "Space Bunny" — on Hugging Face and ModelScope.
- MoE architecture: 428B total parameters, 23B active per token
- Reported scores: 73.7% on DeepSWE v1.1 and 61.7 on OSWorld V2
- SGLang and hosts like Together AI supported it on day one
- Author's tip: to understand MoE inference, start from the model card and work out the memory math — all 428B parameters must fit in memory, though each token only routes through 23B
Related event: MiniMax Open-Sources M3.1-Flash, a 428B-Parameter Multimodal MoE Model(16 posts)→
More from Models
- K3's edge is scale: whether GLM and DeepSeek can train truly big models may doom Moonshot — teortaxesTex · 2026-10-24
- Kimi K3.1 Pro jumps to #8 on WebDev Arena, matching Claude Opus 5 at 70% lower cost — testingcatalog · 2026-10-24
- Every model is benchmark-optimized, says investor: public leaderboards are a waste of time — bindureddy · 2026-10-24
- Users push back on coding-agent weekly limits: resets come too early to use the quota — umeshonai · 2026-10-24
- Moonshot says K3.1 Pro's offensive cyber capability is deliberately capped below frontier — dejavucoder · 2026-10-24
- Researcher links recent lone-actor attacks to abliterated DeepSeek 4.1 Flash via ARTEX — MatthewMcAteer0 · 2026-10-24