Moonshot AI Open-Sources MoonEP: A Perfectly Balanced Expert Parallelism Library for MoE Training

Moonshot AI open-sourced MoonEP, an MIT-licensed expert-parallel library for Mixture-of-Experts training.

MoonEP uses online-planned redundant experts, fused permute/unpermute, and static buffer shapes to keep loads balanced across ranks.

The company says MoonEP is one of the components behind Kimi K3’s claimed 2.5x scaling-efficiency gain.

MoonEP was released alongside Kimi K3 Open Day materials and two other infrastructure codebases.