Kimi K3 debuts as first open 3T-class model; full weights land July 27

Moonshot’s Kimi team unveiled Kimi K3, a 2.8-trillion-parameter model it bills as the world’s first open model in the 3-trillion-parameter class. The architecture pairs two new components—Kimi Delta Attention and Attention Residuals—with native vision and a 1-million-token context window, and pushes Mixture-of-Experts sparsity to activate just 16 of 896 experts via a Stable LatentMoE framework. The company claims roughly 2.5x better scaling efficiency than Kimi K2. K3 is live now across Kimi.com, Kimi Work, Kimi Code, and the API, running max-effort reasoning by default, with open weights promised by July 27, 2026 and a technical report to follow.

The pitch centers on long-horizon, agentic work rather than raw benchmark supremacy—Kimi concedes K3 still trails proprietary leaders Claude Fable 5 and GPT 5.6 Sol, while claiming it beats every other model it tested. The showcased results lean heavily on coding and research autonomy: competitive GPU kernel optimization (K3 reportedly handled most of its own team’s kernel work late in development), a from-scratch Triton-like compiler called MiniTriton that rivals Triton and torch.compile, a 48-hour autonomous run that designed a working chip on open EDA tools, and a two-hour reproduction of astrophysics results that would normally take a researcher a week or two.

Beyond code, Kimi is positioning K3 as a knowledge-work engine, with interactive research dashboards built from thousands of web fetches, plus new Widgets and Dashboard features in Kimi Work for persistent, visual outputs. As with all such vendor announcements, the eye-catching case studies are self-reported and not independently verified; the real test will come when the weights and evaluation details ship publicly later this month.