Moonshot AI launches Kimi K3, a 2.8-trillion-parameter open-weight model, atop the coding leaderboard

Moonshot AI launched Kimi K3 on July 16, 2026, a 2.8-trillion-parameter open-weight multimodal model that debuted at No. 1 on LMArena’s Frontend Code Arena with a 1,679 Elo score.

The Beijing-based startup, backed by Alibaba, calls K3 the largest open-weight model yet shipped and is aiming it squarely at the coding and agent workloads where U.S. labs have led. On that arena, K3’s 1,679 score edged Anthropic’s Claude Fable 5 at 1,631 and OpenAI’s GPT-5.6 Sol at 1,618. Moonshot positions K3 as competitive with the strongest proprietary systems on coding and agentic benchmarks, though it acknowledges its overall performance still trails Claude Fable 5 and GPT-5.6 Sol — a self-assessment independent testers have not yet reproduced.

The model is built on a Kimi Delta Attention and Attention Residuals architecture, with a Stable LatentMoE framework that activates 16 of 896 experts and a 1-million-token context window.

Kimi K3 is available now through Kimi.com, Kimi Code and the Kimi API, and Moonshot said it will publish full model weights on Hugging Face by July 27, 2026, according to its launch post. That open-weight release is the part to watch: Moonshot’s own benchmark comparisons stay vendor-reported until third parties run the model, but downloadable weights let anyone test the coding claims directly.

The debut lands as investors reassess how far China’s open-weight push can undercut U.S. chip and model spending. A single arena placement, even a top one, measures front-end code generation, not the broad reasoning or safety profile enterprises weigh before deploying a frontier model.