Moonshot AI officially open-sourced its newest large language model, Kimi K3, on July 28, releasing the model weights, a full technical report, and several core training infrastructure tools—including the high‑performance communication library MoonEP, the attention operator FlashKDA, and the distributed agent sandbox system AgentEnv.
With 2.8 trillion parameters, Kimi K3 is the company's most powerful model to date. Built on a mixture‑of‑experts (MoE) architecture, it natively handles visual understanding and supports a 1‑million‑token context window. While its parameter count is roughly three times that of its predecessor, Kimi K2.5, innovations like Kimi Delta Attention, Attention Residuals, and MoonEP have boosted intelligence output per unit of compute by about 2.5×.
The company says the Kimi K3 weights are free to use for internal R&D and product integration, including commercial applications.
The three accompanying infrastructure releases—MoonEP, FlashKDA, and AgentEnv—span the full pipeline from high‑performance communication and operators to distributed reinforcement‑learning environments.
Moonshot AI said the open‑source move is aimed at lowering the barrier to AI access, spurring innovation, and giving users stronger data control, privacy, and ownership.