AI Glossary / MoE (Mixture of Experts)

🏗️ Models and training

Mixture of Experts is an architecture where the model consists of many specialized subnetworks ("experts"), but activates only a few of them for each query. This lets the model have a huge total parameter count yet still run fast and cheap - because only a slice works at once. MoE powers many leading models of recent years, especially open ones. For users it's a technical detail, but it explains an apparent paradox: how models with hundreds of billions of parameters can respond instantly.