KimiAI releases KimiK3 model weights and technical report

On July 27, KimiAI released the KimiK3 model weights and technical report. The model adopts a 2.8 trillion parameter Mixed Expert (MoE) architecture, supports native visual understanding capabilities, and has a 1 million Token context window. Officials said that KimiK3 adopts a new model architecture, which achieves an intelligent improvement of about 2.5 times per unit of computing capacity compared to simply increasing the parameter scale. At the same time, the Dark Side of the Moon has also opened up related technology stacks, including a high-performance Attention core, a MoE communication library, and infrastructure to support large-scale operation of agent environments.