← Back to Glossary
Mixture of Experts (MoE)
Phase 02 · LLM Fundamentalscore-concept

An LLM architecture that routes each token to a small subset of specialized sub-networks ('experts') instead of running the whole model.

Why it exists

Allows scaling model parameter capacity to hundreds of billions without increasing compute cost per token.

How it works

A routing gate network evaluates incoming tokens and routes them to 2 out of 8 expert layers (e.g. Mixtral 8x7B), giving huge capacity at fast inference speeds.

Where you'll use this