Mixture of experts (MoE)

A model architecture in which only some specialised sub-networks (“experts”) are active for each request.

An MoE model consists of many expert networks and a router that decides for each token which experts are used. A model can thus have a very large total number of parameters but uses only a small fraction per step.

This makes large models faster and cheaper at inference. Many current frontier and open-weight models use this architecture.