MiniMax published the weights of MiniMax-M3.1-Flash on Hugging Face on Thursday, keeping the open-release date it set when the model launched on its API on October 19 and turning its fastest M-series model into one that anyone can download, inspect and self-host. The repository went live with the full model card, the community licence and serving recipes — the step that turns an API product into part of the open-weight ecosystem.
What Shipped In The Weights Release
M3.1-Flash is a native multimodal mixture-of-experts model of roughly 428 billion total parameters, of which about 23 billion are activated per token. It takes text, image and video inputs, produces text, and supports a context window of up to one million tokens. Reasoning is always on, with effort levels from low to max.
Three changes separate it from June’s M3. Sparse attention now runs throughout the model — MiniMax Sparse Attention replaces the first three full-attention layers. Quantization-aware training is applied to the routed experts and the attention paths during training, shrinking the serving footprint. And DSpark speculative decoding, with its draft head included in the release, replaces the earlier EAGLE-like multi-token approach. MiniMax recommends temperature 1.0 and top_p 0.95, and ships day-one serving support through SGLang and vLLM recipes, TokenSpeed, and FlagOS multi-chip deployment spanning Chinese silicon alongside NVIDIA and AMD hardware.
The Benchmarks MiniMax Claims
The model card leads with agentic and coding scores: 73.7% on DeepSWE v1.1, 76.2 on BankerToolBench, 61.7 on OSWorld V2, 82.2 on CyberGym and 29.6 on Agents’ Last Exam. The Decoder’s launch coverage of the API version put its DeepSWE cost at under $0.30 per task, with the hosted model priced at $0.15 per million input tokens and $0.60 per million output tokens.
The natural comparison is DeepSeek’s V4.1-Flash, the rival Chinese flash model released days earlier. On The Decoder’s head-to-head, DeepSeek leads on the Intelligence Index, 45 to 42, and edges CritPt 24% to 22%, while MiniMax pulls ahead on Humanity’s Last Exam, 48% to 43%, and on AA-Omniscience factual accuracy, 56% to 43%. MiniMax also came out slightly cheaper per task — about $0.22 against $0.26 — because it needs fewer output tokens. Both models are now open-weight; DeepSeek’s V4.1-Flash shipped its weights and a search API last week.
The Licence Terms
The weights ship under the MiniMax Community License, which we read in full on the repository. Non-commercial use is free. Commercial use requires a prominent “Built with MiniMax M3.1-Flash” credit on the product’s website, interface or documentation; products and services generating more than $20 million in yearly revenue must obtain prior written authorization from MiniMax, while smaller commercial users need only send a one-time notice. Military use is prohibited outright, as are uses aimed at exploiting minors or spreading harmful disinformation.
Why It Matters
M3.1 Flash’s weights release matters because open weights are where the Chinese labs are winning the self-hosting market, and this is now the strongest card in that hand: a frontier-adjacent spec sheet, the cheapest API pricing in its class for those who stay hosted, and a downloadable model for those who do not. Every open release of this calibre pressures the API-only strategy of the US frontier labs and gives enterprises a credible exit from per-token pricing. The licence tells its own story — a $20 million revenue threshold and a military ban mean “open” still comes with strings, a middle path between fully open weights and a closed API, and one that regulators in Washington and Brussels will read as closely as developers do. Our explainer on the open-weight versus closed fight sets out what is at stake in that distinction.
Background
MiniMax launched M3.1 Flash on its API on October 19, revealing it as the full version of the model that had topped leaderboards anonymously as “Space Bunny” — our launch coverage noted the weights were promised for October 23, a date The Decoder and SiliconANGLE also reported. M3’s weights have been on Hugging Face since June under the same community licence, and the Flash line is MiniMax’s bid for the high-volume end of the market, where its own figures claim more than a trillion tokens processed by real users in the model’s first 72 hours on the API.
0 Comments