On July 28, 2026, Kimi is set to roll out the highly anticipated K3 Open Day event. At this event, the model weights and a comprehensive technical report for the Kimi K3 will be unveiled. Additionally, three crucial infrastructure technologies that underpin its training—MoonEP, FlashKDA, and AgentEnv—will be made open-source. Kimi K3 stands out as a Mixture-of-Experts (MoE) model boasting a staggering 2.8 trillion parameters. It comes equipped with native visual understanding capabilities and can support a context window of up to a million tokens. To put things in perspective, its parameter scale is roughly three times that of its forerunner, Kimi K2.5. Thanks to a slew of technological breakthroughs, Kimi K3 has achieved a 2.5-fold increase in scaling efficiency.

Users are welcome to download and deploy the model for their in-house R&D endeavors or to integrate it into their products. The technical report delves deep into the core technologies that power Kimi K3, such as KDA+AttnRes, Stable LatentMoE, and MoonViT-V2. It also sheds light on the post-training and evaluation procedures. The trio of Infra technologies, on the other hand, provide robust support in areas like high-performance communication, high-performance operators, and distributed RL environments. It's worth noting that FlashKDA has already been open-sourced in the past. However, MoonEP and AgentEnv will make their official open-source debut at this event.

By open-sourcing these remarkable achievements, Kimi is on a mission to make intelligence more accessible, spur innovation, and propel AGI research forward. At the same time, it's also sending a clear message to the market about the potential risks involved and the importance of investing with caution.