Kimi officially opened the Kimi K3 model weight and published the technical report

On 27 July, Kimi officially opened the weight of the Kimi K3 model and published the technical report. The K3 model uses a structure of 2.8 trillion composite parametric experts (MoE) to support native visual understanding and has a 1 million Token context window. Officially, Kimi K3 introduced a new model structure that achieved about 2.5 times more intelligent growth in the unit of calculation than simply increasing the size of the parameters. At the same time, a technology warehouse has been opened, including the high-performance Attention kernel, the MoE communications library and infrastructure to support large-scale operation of the intelligent environment。