Kimi K3 open-sources multiple core components to support large-scale agent training

PANews, July 27: While the Kimi.ai team opened up the Kimi K3 model weights and technical report, it has continued to open-source multiple underlying infrastructure components, including MoonEP, a high-performance communication library for distributed MoE training; AgentENV, a distributed environment system for large-scale agent workflows (in cooperation with kvcache-ai); and FlashKDA, a high-performance kernel Kimi Delta Attention based on CUTLASS. The official says these components can reduce communication and inference overhead in large-scale MoE and agent reinforcement learning training, and can be used as a plug-and-play backend for flash-linear-attention.