GenAIWiki
Model Architecture

Mixture of Experts

A mixture-of-experts model routes each token through a selected subset of specialized neural-network components instead of activating the full parameter set.

Expanded definition

Mixture of experts, or MoE, is a neural-network architecture with multiple expert components and a router that chooses which experts process each token. Sparse activation can increase total model capacity without using every parameter for every token. Total parameters and active parameters therefore describe different things. A model may have a very large total parameter count while activating a smaller subset during inference. This can improve compute efficiency, but it does not make serving automatically cheap or simple: weights still need storage and memory, routing must be balanced, and distributed communication can become a bottleneck. MoE is an architecture choice, not a quality guarantee. Teams should evaluate the exact model on task quality, latency, throughput, memory, and serving reliability.

Mixture of experts vs dense model

DimensionMixture of expertsDense model
Token pathRouter selects a subset of expertsEach token uses the same dense blocks
Parameter reportingTotal and active parameters differMost model parameters participate in each pass
Serving challengeRouting, memory placement, and communicationLarge matrix compute and memory bandwidth
QualityDepends on training and routing qualityDepends on training and architecture quality

MoE can provide more total capacity per active-token compute, but production value depends on the exact model and serving stack.

Related terms

Explore adjacent ideas in the knowledge graph.

Mixture of Experts FAQ

What is Mixture of Experts?

A mixture-of-experts model routes each token through a selected subset of specialized neural-network components instead of activating the full parameter set.

How is Mixture of Experts used in AI systems?

Mixture of experts, or MoE, is a neural-network architecture with multiple expert components and a router that chooses which experts process each token. Sparse activation can increase total model capacity without using every parameter for every token. Total parameters and active parameters therefore describe different things. A model may have a very large total parameter count while activating a...

Related

Comparisons, tools, and models that connect to this idea.