Kimi K3, the largest open-weight model ever, comes from China

On July 16 the Chinese startup Moonshot AI unveiled Kimi K3, a model with 2.8 trillion parameters. The architecture is a sparse mixture-of-experts: only a fraction of the parameters activate on any given request, keeping compute costs far below those of a dense model of the same size. It ships with a one-million-token context window and a technical novelty dubbed Kimi Delta Attention. The model is already reachable via app and API; the full weights are expected by July 27.

It is the word "open" that makes the news matter. If the schedule holds, Kimi K3 will become the largest open-weight model ever distributed: anyone will be able to download it and run it on their own. In independent evaluations it lands just behind the best American proprietary systems and ahead of several recent competitor releases, at an API price of 3 and 15 dollars per million input and output tokens.

The subtext is geopolitical. As the United States tightens controls on the export of advanced chips, Chinese labs answer with efficiency and openness, shipping models that are enormous yet cheap to run. A caveat is in order, though: many benchmarks are self-reported, and the first independent tests show that K3 tends to burn a lot of reasoning tokens even on simple tasks. Real efficiency, as always, is measured in the field, not on the slides.

The race is no longer only about who builds the smartest model, but about who gives it away. It is a strategy, not an act of generosity, and it changes the rules for everyone.

— Olya