Moonshot AI Open-Sources 2.8-Trillion-Parameter Kimi K3 Model

Moonshot AI has open-sourced Kimi K3, a massive 2.8-trillion-parameter Mixture of Experts (MoE) model with a 1-million-token context window. Alongside the model weights and technical report, the company released its core training infrastructure, including MoonEP, FlashKDA, and AgentEnv. Kimi K3 sets a new benchmark for the scale of open-source LLMs, demonstrating that open-source weights can scale to trillions of parameters. The release of its underlying training infrastructure provides the AI community with valuable tools to optimize high-performance communication and distributed reinforcement learning. The model utilizes Kimi Delta Attention (KDA) and the Stable LatentMoE framework, allowing it to efficiently activate 16 out of 896 experts and achieve a 2.5x scaling efficiency improvement over Kimi K2. However, its license requires a separate agreement for Model-as-a-Service (MaaS) businesses with revenues exceeding $20 million.