MiniMax M3.1 Flash open weights on Hugging Face (~428B/23B MoE, 1M ctx); vLLM and Together day-0

Why it matters

【Key points】

On Oct 23 (Beijing morning) MiniMax shipped the open weights promised in e/957 (“Model weights will be released later”). HF MiniMaxAI/MiniMax-M3.1-Flash landed as one Init with ~48 shards ~02:26 UTC (10:26 CST). @MiniMax_AI posted ~01:06 UTC and named day-0 partners.

Card: multimodal MoE, ~428B / 23B active; text/image/video in, text out; 1M context; MSA throughout; DSpark draft head; QAT. Vendor scores: BankerToolBench 76.2, DeepSWE v1.1 73.7%, OSWorld V2 61.7, ALE 29.6, CyberGym 82.2.

License is MiniMax Community License (not MIT): free non-commercial; commercial use needs “Built with MiniMax M3.1-Flash”; >~$20M revenue needs prior written OK via api@minimax.io, else a one-time notice. Military use banned in appendix. Same day: vLLM Day-0, Together, SGLang, NeMo, Inferact→vLLM, Tenstorrent; mlx 3/4-bit quants.

【Verifiable】

https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash

Repo LICENSE (MiniMax Community License)

https://x.com/MiniMax_AI/status/2113436864904610008

https://vllm.ai/blog/2026-10-23-minimax-m31-flash

https://tenstorrent.com/newsroom/minimax-m31-flash-now-live-on-tenstorrent

e/957 promise: https://www.minimax.io/blog/minimax-m31f

【Our read】

1. After e/952 (docs GA) and e/957 (blog, price, Space Bunny) left weights “later,” the HF drop completes self-host—not another leaderboard splash. It tracks the China Flash open-weight line readers already follow via DeepSeek V4.1 Flash.

2. Card seats it next to DeepSeek V4.1 Flash, Kimi K3, GLM-5.3, GPT-5.6 Sol, Opus 5. On MiniMax’s table, agent scores near closed peers; Terminal Bench / NL2Repo still trail Opus. Attribution + >$20M gate = open weights with commercial guardrails, not Apache/MIT.

3. Day-0 vLLM/SGLang/Together/NeMo/Tenstorrent shortens HF→runnable for private/1M agent stacks. e/957 pay-go ($0.15/$0.60) still splits API vs self-host.

4. Limits: vendor benches; Pro undated; military ban and revenue gate need legal review. mlx quants are convenience, not primary deliverable.