What is MiniMax M3, and what does its community licence require?

MiniMax, a Shanghai AI company, released the weights of its M3 model on Hugging Face yesterday. Weights are the numbers a model learned in training, and having them lets you run the model on your own machines. The hosted version launched on June 1, when the company promised files and a technical report within ten days. They came eleven days later, close to that promise.

M3 is built to read very long inputs at lower cost. It also comes with one of the more detailed licences among the open models, and that licence deserves a careful read before any commercial use.

What was released

The model card describes M3 this way:

A parameter is one learned number, and a token is a small piece of text. M3 is a mixture-of-experts model, so only part of it works on each token: compute depends on the 23 billion, and memory on the 428 billion.

The reasoning modes are worth noticing. Several rival models always think before answering, whereas M3 lets you switch that off to cut delay and cost, or lets the model decide.

How sparse attention works

The new idea is MiniMax Sparse Attention (MSA). In a language model, attention is the step where the model looks back over everything so far to decide what matters for the next word. Normally, this gets more expensive as the input grows, and at a million tokens the memory needed to keep that history becomes the main cost.

Think of reading a very long legal file. You would not reread every page before writing each sentence. You'd check the index, pull the few relevant sections and read those closely. MSA works in a similar spirit. Implicator's summary of MiniMax's description says a pre-filtering step picks the relevant blocks before the full attention calculation runs.

MiniMax reports speedups against M2 at one million tokens of 9 times for prefill (reading the input) and 15 times for decode (writing the output), and says compute per token falls to one-twentieth. The company has published the operator and a technical report. Those are its own measurements, so independent tests will matter more.

What it takes to run

The arithmetic is simple. Take the 444 GB version and divide by 80 GB, a common graphics-card size for data centres. That means at least six cards for the weights alone, plus working memory for your context. The vLLM setup guide centres on eight-GPU machines. MiniMax also lists SGLang, Transformers, KTransformers, Unsloth and AMD's ATOM.

A million tokens of context does not come free either. A long session holds a great deal of data in memory, and the benefit of MSA will depend on your engine supporting it well.

As a result, most people will use a hosted version first. Implicator lists MiniMax's standard prices at $0.60 US per million input tokens and $2.40 per million output tokens.

Do you have to contact MiniMax to use M3 in a business?

Yes, in nearly every commercial case, though usually only by a short email. The grant at the top covers non-commercial purposes. Commercial use is then allowed on conditions:

"Commercial Use" is drawn widely. It includes selling products that rely on M3, commercial use of its APIs, and deploying a fine-tuned version for any commercial purpose. Note that the revenue test applies to the products and services involved, as the licence words it, which differs from the group-wide tests some other labs use.

An appendix bans several uses, including anything supporting a military purpose, harming minors, and generating false information intended to cause harm. Those bans apply to everyone and to derivatives, so a company in defence supply chains should stop and read before it goes further.

The practical point is that this is not Apache 2.0 or MIT, and "open-weight" should not be taken to mean "free for any business without contact". The steps are light for a small firm: add a line of attribution and send one email. Keep a copy of that notice in your records.