Kimi AI announced a new model architecture that it says increases intelligence per unit of compute by 2.5x, not merely through parameter scaling. Alongside Kimi K3, the company will open supporting components including high-performance attention compute kernels, an MoE communication library and infrastructure for large-scale agent environments.

2026-07-27

Kimi AI announced a new model architecture that it says increases intelligence per unit of compute by 2.5x, not merely through parameter scaling. Alongside Kimi K3, the company will open supporting components including high-performance attention compute kernels, an MoE communication library and infrastructure for large-scale agent environments.