MiniMax Posts Open Weights for M3.1 Flash, a 428B Agent Model Under a Licence With a $20M Revenue Line

MiniMax has released open weights for M3.1 Flash, the agent and coding model it launched as an API on October 19. The files went up on Hugging Face and ModelScope on October 23. M3.1 Flash is a multimodal mixture-of-experts model with about 428 billion parameters in total, of which 23 billion are active for each token. It accepts text, images and video, writes text, and handles up to 1 million tokens of context.

The weights are open, but the licence is not a standard open-source one. They ship under the MiniMax Community License, and the terms matter before anyone builds a product on them.

What the licence allows

The MiniMax Community License grants free use, copying, modification and redistribution for non-commercial purposes. Commercial use is allowed with conditions:

• Any commercial product or service built on the model or a derivative must prominently display “Built with MiniMax M3.1-Flash” on a related website, interface, blog post, about page or product documentation.
• If those products earn more than $20 million a year in revenue, the operator needs separate, prior written authorization from MiniMax.
• Below that line, commercial users still owe MiniMax a one-time notice by email.

The licence defines commercial use broadly. It covers paid products and services that rely on the model, commercial use of APIs built on it, and commercial deployment of fine-tuned or otherwise modified versions. An appendix of prohibited uses bans, among other things, any military purpose, harm to minors, and generating false information intended to cause harm.

Built for agents and long tasks

MiniMax pitches M3.1 Flash at coding, agentic workflows, computer use and cybersecurity. It builds on M3 with three changes. Sparse attention now runs through the whole model, replacing the first three full-attention layers. The routed experts and attention paths were quantized during training to cut memory and serving cost. And a speculative-decoding method called DSpark replaces M3’s earlier multi-token prediction approach; the release includes its draft head.

When the API launched on October 19, MiniMax said an earlier version of the model had been running anonymously under the name “Space Bunny.” It also said M3.1 Flash costs a third as much as M3 and runs three times as fast. Hosted pricing is a flat $0.15 per million input tokens and $0.60 per million output tokens, with cached input at $0.015.

MiniMax’s benchmark table

Every score below comes from MiniMax’s model card and was run or collected by MiniMax, often inside third-party agent harnesses such as Claude Code, Codex, OpenCode and mini-SWE-agent.

On MiniMax’s numbers, M3.1 Flash scores 62.8 on Automation-Bench, ahead of every model in its table, and 76.2 on BankerToolBench, level with Opus 5. It scores 73.7% on DeepSWE v1.1, 61.7 on OSWorld 2.0, 29.6 on Agents’ Last Exam and 82.2 on CyberGym.

The same table shows where it trails. Opus 5 leads on Terminal Bench 4.0 (51.8 against 38.1), NL2Repo-Bench (75.3 against 61.2) and OSWorld 2.0 (70.6). DeepSeek’s V4.1 Flash-1017 checkpoint beats M3.1 Flash on Terminal Bench (42.3) and DeepSWE (77.0). On Humanity’s Last Exam without tools, M3.1 Flash’s 43.4 sits behind GPT-5.6 Sol at 49.5 and Opus 5 at 55.6. MiniMax notes that its results are for the official weights and that third-party quantizations may differ.

Running it yourself

The model can be downloaded with the Hugging Face CLI. MiniMax recommends SGLang, vLLM or TokenSpeed for serving, says Transformers, KTransformers, Unsloth and ATOM support are in progress, and points to FlagOS for multi-chip deployments. Its recommended sampling settings are a temperature of 1.0 and top-p of 0.95, and reasoning effort can be set from low to max in the chat template.

At 428 billion parameters, this is a multi-GPU model, and MiniMax notes that usable context and throughput depend on GPU memory and serving setup. For teams that only need to call the model, the hosted API is the simpler route. The download matters most to those who want to fine-tune it, keep data on their own hardware or serve it at scale, which is where the licence’s revenue line and notice rules apply.