Moonshot releases 2.8T-parameter Kimi K3
Moonshot's Kimi K3, a 2.8-trillion-parameter model with a 1M-token context, is now the largest open-weight model, though the company admits it still trails Claude Fable 5 and GPT 5.6 Sol.
Moonshot AI introduced Kimi K3, a 2.8-trillion-parameter model with native vision and a 1-million-token context window. The company calls it the first open "3T-class" model. It is aimed at long-horizon coding, knowledge work and reasoning.
K3 is available now on Kimi.com, Kimi Work, Kimi Code and the Kimi API. At launch it runs at max thinking effort by default, and low- and high-effort modes will come in later updates. Moonshot says the full weights will be released by July 27, 2026. A technical report covering architecture, training and evaluations will come out at the same time.
The model is built on two of Moonshot's own architecture changes, Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). It also uses a sparser Mixture-of-Experts design that activates 16 of 896 experts, paired with a framework the company calls Stable LatentMoE. Moonshot puts the gain in overall scaling efficiency at about 2.5× over Kimi K2.
The company says K3 still trails Claude Fable 5 and GPT 5.6 Sol overall, but beat the other models it tested. Its case studies include:
- GPU kernel optimization: given up to 24 hours per run, K3 performed competitively with Fable 5 and well ahead of Opus 4.8, GPT 5.6 Sol and GPT 5.5. Moonshot notes that Fable 5 was evaluated by a third party and its results may include fallback behavior.
- Compiler: K3 built MiniTriton, a Triton-like compiler that matched or beat Triton and torch.compile on some benchmarks.
- Chip design: in a 48-hour autonomous run, K3 produced a 4 mm² design on the Nangate 45nm library that closes timing at 100 MHz and decodes over 8,700 tokens/s in simulation.
All results are Moonshot's own, and the full benchmark tables and pricing were not included in the material reviewed.