Moonshot released weights and a technical report for Kimi K3, a 2.8 trillion-parameter Mixture-of-Experts model featuring a 1 million-token context window and native vision capabilities. The architecture delivers 2.5 times more intelligence per unit of compute, utilizing a dynamic setup where only 16 of its 896 experts activate per token.

The release bundles the model with open-source infrastructure tools for training and inference, including a high-performance attention kernel, a communication library for expert-parallel workloads, and a distributed system for scaling agent environments. Commercial deployment is governed by a modified MIT license that requires companies earning over $20 million annually to secure a separate commercial agreement, while products exceeding 100 million users or $20 million in monthly revenue must display the model name. The weights are currently available for public download.