Moonshot published Kimi K3 open weights. The license carries the model's own name.

Moonshot AI published downloadable weights for Kimi K3 this week: a mixture-of-experts model with 2.8 trillion total parameters, 104 billion of them active per token, 896 experts with 16 routed per token, and a context window of 1,048,576 tokens. The model card lists the license as the Kimi K3 License. That last line deserves more reading time than the parameter count.

Total parameters set the storage bill. Active parameters set the serving bill. A 2.8T checkpoint has to be held somewhere, which means multi-node hosting and a memory footprint measured in terabytes at full precision, but only 104 billion parameters fire per token, near 3.7 percent of the model. That ratio is the reason a model this size can be served at all, and it is the number to quote back at anyone stacking 2.8T against a dense model's parameter count as though the two describe the same amount of work per token.

Sparse routing moves the bottleneck rather than removing it. 16 experts of 896 per token keeps arithmetic down and leaves the whole checkpoint resident in memory, so the constraint shifts from compute to interconnect and capacity planning.

The download is a planning item on its own. One byte per parameter at 8-bit quantization puts the checkpoint near 2.8 terabytes, and 16-bit weights double that before activation memory, KV cache or any serving overhead enters the arithmetic. Teams that have never moved a multi-terabyte artifact into their own infrastructure should price the transfer, the storage tier and the checksum pass as line items with hours attached.

One detail the size coverage skips: Hugging Face lists the model under image-text-to-text, meaning it accepts images alongside text. Multimodal at this scale turns the deployment question from whether the model can be served into whether it can be served for the request mix a team actually has, because image tokens land in the same context budget as everything else.

The license being named after the model is the tell. Open weights and open source are separate claims, and a vendor-specific license means the terms are whatever the document says: field-of-use limits, attribution requirements and conditions on offering the model as a hosted service are all normal in this category. Anyone putting Kimi K3 behind a paying customer reads the license before the evaluation table. Anyone planning to fine-tune reads the clause covering derivative weights.

A one-million-token context window is a capability claim, not a measurement. The figure describes what the model accepts, not what it uses well, and the two diverge in ways that only appear under a named harness running a named task. Retrieval at 900k tokens is a different exercise from summarization at 900k tokens. Teams that need long context should test their own document shapes before budgeting for them, because that context gets paid for per request, at attention scale, in production.

The strategic read is duller than the specifications and matters more. Labs outside the United States spent 2026 releasing frontier-scale weights under their own licenses while the American frontier stayed behind APIs. The outcome is specific: the ceiling on what a company can run inside its own network keeps rising, and it rises on terms written by whoever published the weights. Data-sovereignty arguments that were theoretical two years ago now have a checkpoint attached to them.

The builder's read: if an architecture assumes the strongest available model is an API call, price the alternative once, now, with real numbers rather than instinct. Multi-node inference for a 2.8T checkpoint is a capital and staffing commitment instead of a weekend, and for most teams the hosted API stays cheaper per token by a wide margin. The teams for whom that math breaks are the ones holding data that cannot leave the building, and those teams should spend this week with the license, not the benchmark chart.