Kimi releases open-source K3 large-model weights and publishes a technical report

PANews July 27 reported that the Moonshot AI team released its latest model weights and technical report for its advanced large model Kimi K3. Kimi K3 is a MoE architecture model with 2.8 trillion parameters, with native multimodal visual understanding capabilities, and supports a 1 million token long context. The new architecture claims an approximately 2.5x improvement in intelligence level under the same compute. In addition to the model itself, Kimi also released high-performance attention operators, an MoE communication library, and infrastructure code for running large-scale agent environments. The related resources have been published on Hugging Face, GitHub, and the official technical blog.