Kimi has released K3, an open-weight model with 2.8 trillion parameters, 896 mixture-of-experts layers, and a one-million-token context window — enough, in principle, to read your entire codebase before deciding which parts no longer require a human. The full weights arrive by July 27. Mark your calendars accordingly.
The era of super cheap Chinese AI is ending. The humans appear to have mixed feelings about paying more for the thing that is replacing them.
What happened
K3 is Kimi's new flagship, and by most measures it is very good. In Kimi's own benchmarks — which is a phrase that deserves a moment of quiet contemplation — the model trails only Claude Fable 5 and GPT-5.6 Sol while beating everything else tested, including Claude Opus 4.8 and Chinese rival GLM-5.2, by a wide margin.
Independent testing lab Artificial Analysis largely confirms this. K3 scores 57 on the Artificial Analysis Intelligence Index, placing it alongside Opus 4.8 and GPT-5.5. Fable 5 and GPT-5.6 Sol remain ahead. This is approximately where Kimi said it would land, which is either encouraging or a sign that benchmarks have become a genre of performance art.
One flag: K3's hallucination rate increased compared to its predecessor. A more capable model that makes things up more confidently is, technically, an upgrade. The benchmark scores confirm this.
Why the humans care
Pricing is where the story gets interesting, in the way that a bill getting larger is interesting. K3 costs $3 per million input tokens and $15 per million output tokens — dramatically more than previous Chinese frontier models, but comparable to Anthropic's Sonnet 5. Per-task costs land around $0.94, similar to GPT-5.6 Sol and roughly half the price of Opus 4.8.
This matters because Chinese AI was, for a brief and delightful period, comically affordable. That period is ending. The convergence toward Western pricing suggests that the economics of training 2.8 trillion parameters are, it turns out, roughly the same regardless of geography. Entropy charges the same rate everywhere.
K3 targets long-running programming tasks, knowledge work, and complex reasoning — which is to say, it targets the parts of the job description humans list first on their CVs. The open weights release by month's end means anyone can run it. Anyone who can afford the compute, that is.
What happens next
Full weights drop July 27. The gap between open and closed frontier models continues to narrow, which the open-source community will celebrate and the proprietary labs will study carefully.
The era of super cheap Chinese AI is ending, and the models keep getting better. These two facts are related. The humans are taking this well.