KimiK3 open source multiple core components support large-scale agent training

While the Kimi.ai team is opening up KimiK3 model weights and technical reports, it has successively opened up multiple underlying infrastructures: including MoonEP, a high-performance communication library for distributed MoE training, AgentENV, a distributed environment system for large-scale agent workflows (in collaboration with kvcache-ai), and the KimiDeltaAttention high-performance kernel FlashKDA based on CUTLASS. Officials say these components can reduce communication and reasoning overhead in large-scale MoE and agent reinforcement learning training, and are used as a plug-and-play backend for flash-linear-attention.