The wait lasted one news cycle. A day after the Financial Times reported that Kimi K3 was days away, Moonshot AI shipped it: 2.8 trillion parameters, a mixture-of-experts design that activates 16 of 896 experts per query, a 1-million-token context window, and native image understanding. It is the largest open-weight model ever announced — nearly double DeepSeek’s 1.6 trillion, and roughly seven times the size of Alibaba’s biggest.
The numbers are the story. On GDPval-AA v2 — the benchmark that grades models on real deliverables across 44 occupations and nine industries — K3 scored 1,687: third in the world, behind only Anthropic’s Claude Fable 5 Max (1,815) and OpenAI’s GPT-5.6 Sol Max (1,748), and ahead of Claude Opus 4.8 at 1,600. On AA-Briefcase, a private long-horizon agent test, it took second outright. It ranked first in four of eight task-automation benchmarks, including SpreadsheetBench 2 and BrowseComp — and in blind Arena testing, developers preferred its front-end code to every leading US model’s.
Launch pricing is $3 per million input tokens ($0.30 cached) and $15 per million out. And unlike everything above it on the leaderboard, K3 is scheduled to be free to download on July 27.
The asterisk
“Open,” for now, is a promise. As of this morning there is no public checkpoint, no model card, no technical report, and no safety report — just an API and a date. Moonshot says the weights land July 27 under the modified-MIT-style license it used for earlier Kimi releases. Until they do, K3 is a closed model with an open-source press release. And ten days is a long time in a market where one Bloomberg story about a late model erased $200 billion.
Our take: the leaderboard isn’t the point — the invoice is. The moment those weights go public, every cloud on Earth can host K3 and compete on the price of serving it, which makes third-best-in-the-world the new floor for cheap intelligence. US labs keep pricing power only as long as the gap between first and third stays worth paying for. Moonshot just closed a lot of that gap in one release — while Google’s flagship sits months late.
What to watch
- July 27: do the full weights, model card, and safety report actually ship — and under what license? A slip or a stripped-down release changes this story completely.
- Hosting prices: how fast third-party clouds undercut Moonshot’s own $3/$15 rate once anyone can serve the model.
- Enterprise evals: whether companies that standardized on US models for AI-assisted engineering start benchmarking K3 against their Claude and GPT bills.
- Washington: a frontier-grade open Chinese model is the sharpest argument yet in the export-control and AI-regulation fights already underway.