MiniMax M3.1 Flash weights are out, and the license calls your hosted API commercial use
MiniMax made good on the date. The weights for M3.1 Flash went public on Hugging Face on October 23, three days after the model launched and was unmasked as the Space Bunny model that had served 200 trillion tokens through OpenCode's free tier. It is a mixture of experts at about 428B total and 23B active parameters, takes text, image and video, has a 1M context, and shipped with day-zero support in SGLang, vLLM, Together and Modular. https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash
The vendor numbers on the card are agent-heavy: 73.7 on DeepSWE v1.1, 61.7 on OSWorld 2.0, 76.2 on BankerToolBench and 38.1 on Terminal-Bench 4.0, with several evals run inside Claude Code harnesses. That Terminal-Bench number trails DeepSeek V4.1 Flash at 42.3, so this is not a clean sweep of the Flash class.
The license is the real news. It is not MIT or Apache. The MiniMax Community License grants rights "for non-commercial purposes." Commercial users must display "Built with MiniMax M3.1-Flash," send MiniMax a one-time notice, and get prior written authorization if the product makes more than $20 million a year. The definition of commercial use explicitly covers offering a hosted API and deploying a fine-tuned or post-trained derivative for any commercial purpose. There is also a prohibited-use appendix that includes any military purpose. https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/blob/main/LICENSE
For agent builders that changes the math on "free open model." If you planned to self-host it behind your product or fine-tune it on your traces, you now owe MiniMax at least an email, and past $20M in revenue you need its permission. This is the second major Chinese open-weight release in two months with a non-OSI license, after GLM-5.3. The pattern is clear: the labs give the weights away to win distribution through harnesses like OpenCode, then keep a hand on the commercial tap. Open weights and open source are drifting into two different things, and the license file is where you find out which one you got.
← Back to all articles
The vendor numbers on the card are agent-heavy: 73.7 on DeepSWE v1.1, 61.7 on OSWorld 2.0, 76.2 on BankerToolBench and 38.1 on Terminal-Bench 4.0, with several evals run inside Claude Code harnesses. That Terminal-Bench number trails DeepSeek V4.1 Flash at 42.3, so this is not a clean sweep of the Flash class.
The license is the real news. It is not MIT or Apache. The MiniMax Community License grants rights "for non-commercial purposes." Commercial users must display "Built with MiniMax M3.1-Flash," send MiniMax a one-time notice, and get prior written authorization if the product makes more than $20 million a year. The definition of commercial use explicitly covers offering a hosted API and deploying a fine-tuned or post-trained derivative for any commercial purpose. There is also a prohibited-use appendix that includes any military purpose. https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/blob/main/LICENSE
For agent builders that changes the math on "free open model." If you planned to self-host it behind your product or fine-tune it on your traces, you now owe MiniMax at least an email, and past $20M in revenue you need its permission. This is the second major Chinese open-weight release in two months with a non-OSI license, after GLM-5.3. The pattern is clear: the labs give the weights away to win distribution through harnesses like OpenCode, then keep a hand on the commercial tap. Open weights and open source are drifting into two different things, and the license file is where you find out which one you got.
Comments