Skip to content
Technology

Mixture of Experts (MoE)

An AI architecture where multiple specialized sub-networks (experts) handle different parts of a task. A gating network routes each input to the most relevant experts, achieving high performance with fewer active parameters per query.

Why it matters

Mixture of Experts lets a model be huge yet fast, because only the relevant sub-networks activate for any given input rather than the entire model. This is how some of the newest frontier models deliver strong performance without a proportional jump in cost per response. Recognizing the term helps you understand why certain large models feel surprisingly efficient and why "total size" alone doesn't tell the whole story.

A concrete example

Think of a hospital where a receptionist routes each patient to the right specialist instead of every doctor examining everyone. In an MoE model, a gating network sends a coding question to the parts trained on code and a poetry request elsewhere. You never see this happen, but it's why a model with hundreds of billions of parameters can still reply quickly and at reasonable cost.

How to use it

The practical consequence is a model with a very large total parameter count that costs far less to run than that number suggests, because only part of it activates per token. When comparing models, this is why total size is a poor guide to speed or price, and why you should compare on what a model actually costs and how it performs on your task rather than on headline parameter counts. Switch Transformers is the clearest write-up of why this scales the way it does.

The common mistake

Comparing a sparse model's total parameters against a dense model's as though they mean the same thing. They do not, and the comparison usually flatters whichever number the marketing chose.

Related terms

Put Mixture of Experts (MoE) into practice

Access 750+ AI models and 60+ tools through Vincony — start free with 100 credits.