The dark side of the moon. Kimi releases the MoE architecture basic model K2 and open sources it simultaneously, with a total of 1T parameters.

The dark side of the moon: Kimi has released the MoE architecture basic model K2 and open sourced it simultaneously, with a total of 1T parameters and 32B activation parameters, surpassing other open source models globally in fields such as autonomous programming, tool invocation, and mathematical reasoning. Kimi K2 uses the MuonClip optimizer to efficiently train models with trillions of parameters, and in the context of bottlenecking with high quality data, it improves token efficiency to find new pre-training extension space. K2 has stronger coding capabilities, excelling in general Agent tasks, demonstrating stronger ability generalization and practicality in multiple real scenarios, and the new model is now available online for experience.