MiniMax releases M3.1 Flash, a 428B open-weight MoE under a conditional commercial licence

MiniMax has put the weights for its M3.1 Flash model on Hugging Face, and like most of the interesting open-weight releases this year it comes from a Chinese lab and arrives with a licence worth reading before you build on it. The model itself is a frontier-scale, natively multimodal mixture-of-experts system. The licence is the part that decides whether you can use it for what you have in mind.

What was released

M3.1 Flash is a mixture-of-experts model that the model card describes as roughly 428 billion total parameters with about 23 billion active per token. Its published configuration (60 layers, 128 routed experts with four routed plus one shared expert active per token, a hidden size of 6,144) reconciles to those figures. One wrinkle for anyone checking the page: Hugging Face’s automatic metadata box reports “220B params”, which does not match the architecture and appears to understate the model, so the accurate description is the card’s roughly 428 billion total and 23 billion active. Its context window runs up to one million tokens, with usable length depending on your memory and serving setup.

The weights are present and downloadable without a gate, as safetensors in 8-bit (FP8) formats, and quantised community repackagings for llama.cpp, Ollama and LM Studio appeared within days.

This is server territory, not a laptop model. Even in 8-bit, a model of this size is a multi-GPU or large-unified-memory proposition, with the full download running to a few hundred gigabytes. As with any mixture-of-experts model, the memory you need is set by the total parameter count, not the smaller active count, so the 23-billion active figure describes its speed, not its footprint.

The licence is the story

M3.1 Flash is open weights, not open source. We read the actual licence file, the MiniMax Community License, rather than relying on second-hand summaries, and the terms are specific:

• Non-commercial use is free of charge.
• Commercial use is permitted, with two conditions. You must prominently display “Built with MiniMax M3.1-Flash” somewhere such as your website, interface, or product documentation, and you must send MiniMax a one-time notice.
• Above 20 million US dollars in yearly revenue, commercial use needs separate, prior written authorisation from MiniMax (by email to [email protected]), rather than only a notice.
• The licence prohibits using the model to generate illegal content, to support any military purpose, to exploit or harm minors, to spread misinformation intended to cause harm, or to promote discrimination or hate.

In our catalogue terms that is permitted-conditional, not the clean “permitted” of an Apache 2.0 or MIT release. For an individual or a small team the practical burden is the attribution line and a one-time notice, both easy to meet. The revenue threshold only reaches larger operators, and even then it is a request for authorisation rather than a flat bar. The point is that the conditions exist and differ in the details from other “community” licences, which is exactly why they are worth stating plainly rather than waving through as “open”.

Where it sits

M3.1 Flash joins a crowded field of Chinese open-weight mixture-of-experts models, alongside DeepSeek V4.1 Flash, GLM-5.3 and Kimi K3, that together are a large part of why running capable AI on your own hardware is realistic at all. What is notable here is a multimodal model at this scale that you can genuinely host yourself, if you have the hardware for it.

The licence pattern is becoming familiar: free weights, an attribution string, and a revenue gate. It echoes Meta’s conditional Llama licences and the 20-million-dollar threshold on Meta’s recent Muse Spark Open. The figures and the triggers differ between them, though, and here even modest commercial use carries a notice requirement that a permissive licence would not. Reading the specific terms, rather than assuming one “community licence” matches another, is the habit that keeps you out of trouble.

It also lands against a shifting policy backdrop. Our coverage of China’s draft AI export controls is relevant context: weights already released as open source, which M3.1 Flash now is, sit in the exemption that draft rule carves out, so this release is unaffected by it. The broader theme, that open weights and “freely usable everywhere” are coming apart, is the one to keep in view. For now, M3.1 Flash is a real, downloadable, frontier-scale model, permissive enough to be useful and conditional enough to read carefully. Its full requirements, quantisation sizes and licence conditions are in our catalogue entry.