Moonshot AI put Kimi K3 live this week, and the company is calling it its most capable model to date. It runs 2.8 trillion parameters with a 1 million token context window, native visual understanding, and a new attention setup Moonshot calls Kimi Delta Attention. Details are on the API docs.
The number Moonshot is pushing hardest: 91.2 on BrowseComp, which it reports as state of the art. What makes that claim interesting is the setup. K3 hit it as a single agent, no context compression, no window-management tricks, which is the kind of thing the 1M context is supposed to buy. All self-reported, no independent runs yet.
On knowledge work the picture is more mixed than a clean win. Using Artificial Analysis figures, Moonshot puts K3 at 1687 on GDPval-AA v2. That's behind Claude Fable 5 Max and GPT-5.6 Sol Max, though ahead of Claude Opus 4.8 Max at 1600. So third on that board, not first. On the private AA-Briefcase agentic benchmark it lands second, just past Sol Max.
The model is already on kimi.com and through the API, including on our own DS Lab platform. Moonshot says open weights land by July 27.
Bottom Line
Kimi K3 is the largest open model yet at 2.8 trillion parameters, but its 1687 GDPval score puts it third, behind Fable 5 and Sol Max.
Quick Facts
- 2.8 trillion total parameters (company-stated)
- 1 million token context window
- BrowseComp: 91.2, self-reported SOTA, single-agent
- GDPval-AA v2: 1687, third behind Fable 5 Max and Sol Max
- Open weights promised by July 27, 2026
