Kimi K3 aims to be the first open 3-trillion-parameter model

Moonshot AI released Kimi K3 on July 16, a 2.8 trillion parameter model the company is calling the first "open 3T-class model." The weights are not out yet; Moonshot has promised them for July 27, and for now you can only reach K3 through its website and API. Simon Willison, who ran it through his usual tests, notes that the pricing has jumped sharply: $3 per million input tokens and $15 per million output tokens, well above the $0.95 and $4 that earlier Kimi models charged.

On paper the numbers are strong. Moonshot's self-reported benchmarks put K3 ahead of Claude Opus 4.8 and GPT-5.5 on most tasks, though behind Claude Fable 5 and GPT-5.6 Sol. Artificial Analysis measured an overall Elo of 1547 and found K3 leading the Frontend Code arena. The catch is cost and control. K3 currently offers a single reasoning effort level, "max", so there is no cheaper mode for simpler work. When Willison asked it to draw an SVG pelican on a bicycle, one of his standard checks, the model used 16,658 output tokens, 13,241 of them reasoning, for a single 25-cent image.

The open-weight promise is the part worth watching. A genuinely open model at this scale, if the July 27 release holds, would be the largest weights any lab has published, and Willison's writeup gives a sense of both the ambition and the rough edges.

Why it matters

If you are choosing an open model to build on, K3's size and benchmark claims make it tempting, but wait for the July 27 weights and test the "max"-only pricing against your own workload first, because at $15 per million output tokens it is not a cheap default.