At midnight UTC on July 27, Chinese AI lab Moonshot AI released the full open weights of Kimi K3, a 2.8-trillion-parameter model that is now the largest open-weight AI system ever made publicly available. The model uses a sparse mixture-of-experts architecture with 896 experts, only 16 active per token, and a 1-million-token context window, putting it in the same performance class as top closed systems from OpenAI and Anthropic.
The catch: the weights alone are roughly 1.4 terabytes using MXFP4 quantization. Self-hosting requires a multi-node H100 cluster, meaning most teams will access K3 through inference providers like Kimi's own API rather than downloading and running it themselves. But the release matters anyway. [1]
K3 is the first open model to rival frontier U.S. systems in head-to-head benchmarks. K3 beat Anthropic's Claude Fable 5 in the Frontend Code Arena benchmark a result few expected from an open-weight release. The self-hosting lets organizations process sensitive data without routing it through third-party APIs, a critical advantage in sectors like healthcare and defense. The move lands days before China's World AI Conference and puts fresh pressure on Western labs. [2]
Meta has open-sourced Llama at progressively larger scales, but no U.S. company has released weights at the 2.8-trillion level. For now, a frontier-class model is free to download and the AI race just got a little harder to control.