MiniMax Cut the Price of Intelligence and Built a Toll Booth at $20 Million

Share

This week MiniMax released the open weights for M3.1 Flash, and the headline math is genuinely impressive. Around 428 billion total parameters in a sparse mixture of experts, only 23 billion active per token, a full million token context window, and API pricing that starts at fifteen cents per million input tokens. On paper it is yet another sign that open weights keep getting cheaper and more capable, which has been my story all year. But buried in the announcement is a term that matters more to a founder than any benchmark, and almost none of the coverage flagged it. Read the license and you will find a revenue line set at twenty million dollars a year. That line is the real product, and it tells you everything about how open this release actually is.

What actually shipped

Start with the parts that are unambiguously good. The weights are public and you can run them on your own hardware. It is a native multimodal model, meaning it takes text and images in and produces text out, built for coding, agentic office work, computer use and cybersecurity. MiniMax's own numbers claim a sharp jump over its predecessor on agentic benchmarks, including 76.2 on BankerToolBench, 73.7 on DeepSWE, 61.7 on OSWorld and 82.2 on CyberGym. Treat those as vendor claims, not independent verification, but the direction is believable. The architecture leans on sparse attention, quantization aware training and speculative decoding to make a huge model cheaper to serve. Functionally it is one of the cheapest ways to get a genuinely capable coding and agent model today.

Then look at the download. It is 252 gigabytes. That is not a laptop model. For reference, a single 96 gigabyte card like the one in my workstation does not hold it at full precision, and neither does any consumer machine I know of. The honest way to put this is that it is local AI for people with a serious GPU budget, not for the average founder. The scale gate matters because it quietly narrows who the open part of this release actually serves.

The license is the product

Here is the part everyone is skating past. M3.1 Flash is not Apache 2.0 and it is not MIT. It ships under a MiniMax Community License, and the interesting term is a revenue threshold. Below twenty million dollars in annual revenue you can use it freely as long as you give attribution and send a one time notification. Above twenty million dollars in annual revenue you need prior written authorization from MiniMax before you can keep selling on it.

Read that sentence again, slowly. The model looks free and open, and it is, right up until your product gets big enough that the vendor wants a say in your business. Twenty million dollars is not a rounding error for a startup. It is usually the point right after a product stops being a hobby and starts being a real company, with real customers, real contracts and real code built on those weights. By the time you cross that line you have spent a year and a half baking one model into your entire stack, which is exactly when you are least able to leave.

Cheap price, expensive bill later

The thing I keep turning over is how the pricing and the license fit together. Fifteen cents per million input tokens is not a market price, it is an entry strategy. Combined with open weights, it is designed to make committing feel free. You build on it because it is the rational economic choice today, and the vendor knows that every integration you ship makes you cheaper to keep and more expensive to replace. The toll does not need to be high to bind you. It just needs to exist at the moment you become valuable.

That is a fundamentally different relationship than a vendor charging its full margin up front. A vendor you pay a lot from day one is a vendor you evaluate from day one. A vendor that prices like a friend while holding a switch it can flip later has structured your incentives in its favor without you noticing. Open weights in this shape do not remove vendor dependence. They defer the conversation about it until after you have already built everything, which is the worst time to have it.

The pattern keeps repeating

I have written a version of this warning every time a lab releases weights with a license attached this year, from Kimi's revenue gating to the DeepSeek serving release that only runs on its own chips. The lesson is the same each time and it keeps getting confirmed. Open weights are a distribution move, not a promise of independence. The model itself is commoditizing fast, which is great for the cost of intelligence and bad for anyone who thought owning a model was a moat. The actual moat, and the actual leverage, lives in the license, the data, the workflow and the distribution that sit around the weights.

This is not a criticism of MiniMax. Shipping a powerful model cheap is a reasonable commercial play, and founders who stay under a twenty million dollar business get a genuine bargain. The point is that the bargain is a negotiation, not a gift. Both parties entered it, but only one of them wrote the terms.

The test to apply

Before you build a product on any open weights, ask one question. At what revenue, if any, does this stop being mine? If there is a number attached, no matter how generous, the relationship has a second act you have not negotiated yet. Read the license the way you would read a term sheet, because that is what it is, and the term sheet here says the model is yours until you succeed, and then it is theirs. Cheap intelligence is a real gift for the buyer today. Just do not mistake the gift for ownership, because the vendor making this offer is betting, correctly, that you will. The winning move is to build on cheap open weights while keeping your own data, your own workflow and a plan for the day the toll appears.