Kimi K3 Released: 2.8T MoE Open-Weight Model, Near Frontier Performance

Kimi K3 is Moonshot AI's flagship MoE model with 2.8T parameters, announced on July 16, 2026, with open weights releasing July 27. It ranks second in the Vals AI Index and behind only Claude Fable and GPT-5.6 Sol Max in the Artificial Analysis Intelligence Index, with lower cost, and holds first place in the Frontend Code Arena. The architecture features Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with MoE activating 16/896 experts under the Stable LatentMoE framework, achieving ~2.5x scaling efficiency over Kimi K2. The model cuts the US-China frontier performance gap from 6–9 months to 3–5 months. For developers and enterprises relying on open-source LLMs, Kimi K3 offers unprecedented performance, enabling near-frontier applications without closed APIs. Its cost advantage further lowers barriers. For frontier labs like OpenAI and Anthropic, strong open models may compress margins and slow overall AI investment. Chinese government support for open-source AI at the World AI Conference also reduces policy risk.