On the evening of July 27, Moonshot AI made a significant announcement regarding the Kimi K3 model. The company declared the open-sourcing of the model weights for Kimi K3, the publication of a comprehensive technical report, and the concurrent unveiling of three pivotal infrastructure technologies: MoonEP, FlashKDA, and AgentEnv. This open-source initiative encompasses the model's architecture, training methodologies, and the foundational training system.

Developers now have the opportunity to download the Kimi K3 model for local deployment and engage in secondary development, adhering to the specific terms outlined in the Kimi K3 license. The Kimi K3 model is built upon a Mixture of Experts (MoE) architecture, boasting a staggering total of 2.8 trillion parameters. It supports a context window of up to 1 million Tokens and inherently possesses visual understanding capabilities.

In comparison to its predecessor, the Kimi K2.5, the Kimi K3 has witnessed a substantial increase in parameter scale, approximately tripling in size. Furthermore, the model's scaling efficiency has seen a notable improvement, roughly 2.5 times more efficient. The trio of technologies released alongside the Kimi K3—MoonEP, FlashKDA, and AgentEnv—are tailored for distinct purposes within the realm of large-scale MoE models. MoonEP is designed for efficient communication among large-scale MoE models, FlashKDA enhances high-performance attention computation, and AgentEnv provides a robust training environment for large-scale Agents. These technologies collectively aim to diminish the obstacles associated with model deployment and development, fostering the expansion of the open-weight model ecosystem.