Kimi K3 is free now. You still cannot use it

1.56 terabytes. That is what the largest open AI model in the world weighs, now that Moonshot put the Kimi K3 weights on Hugging Face overnight into 27 July, a day ahead of schedule.

I went and checked what it would take to start the thing. The answer to “can we run this ourselves” is no for practically every mid-sized company.

Which makes the freedom inside the phrase “open weights” something you cannot use. The licence permits you everything, and you can act on none of it.

And this day still matters to all of us.

The calculation nobody does

Open weights are a model’s trained parameters, published for download. Whoever has them may run the model without asking the vendor.

Storage is the smallest part of it: 2.5 to 3 terabytes covering the download, cache and working copies. That costs nothing.

Loading it is the problem. Kimi K3 has 2.8 trillion parameters and every one of them has to sit in GPU memory at the same time. That only 104 billion of them compute per request makes the model fast, not smaller. Spread 1.56 terabytes across accelerators holding 80 gigabytes each and you are at around twenty cards, and since such servers are built as eight-card nodes, three of them with 24 cards in practice. For production Moonshot recommends 64 or more. It has never published a dependable floor, which leaves the weight of the files as the hardest number there is.

That has a price tag. A single H100 runs between two and three and a half dollars per hour at the cheaper cloud providers, so 24 of them come to 35,000 to 60,000 dollars a month if you keep them running. Buying instead, a ready-built server with eight cards costs around 285,000 dollars, so about 855,000 for the three you need, before power, cooling and the person who operates them. The 64 cards Moonshot recommends for production come to eight such servers, so a good two million.

You can work out the break-even yourself. At three dollars per million tokens, the rent on the minimum configuration alone equals twelve to twenty billion tokens a month. No mid-sized company pushes that much text through a model.

What the licence says

Kimi K3 runs under a modified MIT licence. Download, run, modify, use commercially: all permitted.

The modification only applies above 100 million monthly active users or 20 million dollars in annual revenue from the model, at which point a separate agreement is required. The clause targets providers reselling someone else’s model as their own product. For a company simply using it, it is irrelevant.

Which is where freedom on paper and freedom in practice part ways. Everything is permitted to you. None of it is available to you.

The real gain

If you cannot run it yourself, what is the excitement about?

Because open does not mean you have to be the operator. It means you get to choose who is.

A closed model exists at exactly one provider, at their price, in their data centre. With an open one, several operators compete. K3 was bookable at multiple providers within hours, and the competition there is not about who has the better model but who runs it cheaper.

The price, because it always comes up

Through the API, K3 costs three dollars per million input tokens and fifteen per million output at Moonshot. Repeat requests sharing a prefix drop to roughly thirty cents. Claude Opus 5, the other story this week, sits at five and twenty-five.

What you are using there is not the free file but an operator’s service. Exactly as with a closed model, except that this time you choose which operator.

So open and closed models no longer occupy separate price brackets. That turns the choice of model into a question of the task rather than the budget.

What this means for you

Open weights are rarely an invitation to play data centre. At this size they certainly are not.

What they give you is leverage. A model several operators can serve cannot be locked to one contract. And one that can run in Europe does not have to run in the United States just because its maker sits elsewhere. Moonshot sits in Beijing, which is what makes the calculation interesting to some people in the first place.

This is the choice we designed RelationFlow around. The major closed models sit in one place, and anyone wanting to add an open model at a European operator or on their own hardware can do so on any plan. For the cases where genuinely nothing may leave the building, that is the difference between a second isolated tool and one more option in the same workspace.

Why one model is never enough holds in both directions.

The rule I take from this week: when sizing up a model, do not ask first what it can do. Ask how many people can offer it to you. The answer to the second question lasts longer.

Sources: Moonshot on Hugging Face. Hardware pricing from getdeploying and CloudZero, as of July 2026.