The Biggest Open AI Model Ever Is Free. Running It Takes 1.4 Terabytes.

Moonshot AI released the full weights for Kimi K3, a 2.8-trillion-parameter model, for free. Renting it costs exactly what Claude Sonnet 5 costs, which tells you what open weights actually buy: an exit, not a lower bill.

If your company has data that legally cannot leave your network, you have spent the last two years picking between two bad options: send it to an AI vendor anyway and paper over the problem with a contract, or run an open model that was visibly worse at the actual work. As of last night there is a third option, and its price tag is the interesting part. Moonshot AI published the full weights for Kimi K3, a 2.8-trillion-parameter model, for free download. Actually running it takes roughly 1.4 terabytes of fast memory before you load a single line of context. Rent it from Moonshot instead and you pay $3 per million input tokens and $15 per million output, which is exactly what Claude Sonnet 5 lists at.

So the headline is not that frontier AI just became free. It's that frontier AI just became portable, and portable turns out to cost about what renting it does.

What actually shipped

The weights went up at moonshotai/Kimi-K3 on the evening of July 26, about a day ahead of the July 27 date Moonshot had been quoting. The model card is the thing to read, not the coverage.

K3 carries 2.8 trillion total parameters with 104 billion active on any given token. It's a heavily sparse mixture of experts: 896 experts, 16 of them selected per token. The context window is 1,048,576 tokens. And the weights ship in MXFP4, a four-bit format applied through quantization-aware training rather than crushed down after the fact, which is the only reason a model this large fits in 1.4 terabytes at all.

That number is worth sitting with. At sixteen-bit precision the same weights would want something like 5.6 terabytes. Four-bit training gets it to 1.4, and a technical breakdown of the release puts the practical serving floor at roughly 64 GPUs in the 80GB class once you leave headroom for context and activations. Not 64 GPUs to make it fast. 64 GPUs to hold it.

The license is the part that decides whether you can use this

Skip the GPU math for a second, because for most companies it isn't the deciding factor. The license is.

The Kimi K3 License grants broad rights and then draws exactly two lines. If you operate a model-as-a-service business and your revenue crosses 20 million dollars over any consecutive twelve months, you need a separate agreement with Moonshot before serving K3 commercially. And if your product has more than 100 million monthly active users or more than 20 million dollars in monthly revenue, you have to display "Kimi K3" prominently in your interface.

Internal use is exempt from both.

Read that again if you run a normal company. Your support triage, your contract review, your document pipeline, your internal agents: none of it trips either clause, no matter how much revenue you do. The restrictions are pointed at the companies who would resell K3 as an API, not at the ones who would use it to get work done. That is a meaningfully different posture from "open weights, terms and conditions apply," and almost nobody covering this release led with it.

Which gets at what open weights actually buy you, and it is not a smaller invoice. It's an exit. Every AI vendor negotiation you have ever had, you conducted knowing the vendor knew you had nowhere comparable to go. A downloadable 2.8-trillion-parameter model changes the shape of that conversation whether or not you ever download it. Leverage does not require you to exercise the option. It requires the option to exist.

The honest part

Here is what the open-source-wins crowd will skip past.

Free weights are not cheap tokens. Moonshot prices K3 at $3 and $15 per million, the same as Claude Sonnet 5 and cheaper than Claude Opus 5 at $5 and $25, per Anthropic's published pricing. Cached input drops to $0.30 per million, which is real money back if your workload reuses a large prefix. But there's a wrinkle worth knowing before you budget: K3 always reasons at maximum effort. There is no cheaper non-thinking mode, and every reasoning token bills at the $15 output rate. On a classification job or a bulk summarization pass where you do not need deep reasoning, that's a floor you cannot opt out of. The closed vendors will happily sell you a cheaper tier for that work.

And self-hosting does not delete the cost, it relocates it onto your balance sheet and your on-call rotation. A terabyte and a half of GPU memory is a capital line item, a power bill, someone to keep the cluster patched, and an outage that is now yours to own. For a team with genuine infrastructure appetite and steady volume, that math can work out well. For a team without it, the free model is the expensive one the first week it goes down.

The 2.8 trillion parameters are not the story. The story is that "nobody will let us run a model that good on our own terms" stopped being true, and it stopped being true at roughly the price of not bothering.