Kimi K3 Open-Sources Core Infrastructures for Large-Scale Agent Training

Kimi.ai releases K3 weights and open-sources MoonEP, AgentENV, and FlashKDA to optimize MoE training and agent workflows with reduced overhead.

Woofun AI reports that the Kimi.ai team has released the Kimi K3 model weights and technical report alongside a suite of open-source underlying infrastructures. The release includes MoonEP, a high-performance communication library for distributed MoE training, and AgentENV, a distributed environment system for large-scale agent workflows developed in collaboration with kvcache-ai.

Additionally, FlashKDA, a high-performance kernel based on CUTLASS, is now available. These components are designed to reduce communication and inference overhead in large-scale MoE and agent reinforcement learning training, serving as a plug-and-play backend for flash-linear-attention.