Kimi from the Dark Side of the Moon releases the MoE architecture basic model K2 and synchronously opens it up, with a total parameter of 1T

Kimi from the Dark Side of the Moon has released the MoE architecture basic model K2 and simultaneously opened it up, with a total parameter of 1T and an activation parameter of 32B, surpassing other open source models worldwide in areas such as autonomous programming, tool calling, and mathematical reasoning. Kimi K2 uses the MuonClip optimizer to achieve efficient training of trillion parameter models. In the context of high-quality data encountering bottlenecks, it improves token efficiency and finds new pre training expansion space. K2 has stronger coding capabilities and excels in general agent tasks, demonstrating stronger generalization and practicality in multiple practical scenarios. The new model is currently available for open testing. (36Kr)