← All terms
Models
MoE (Mixture of Experts)
A model architecture where only a subset of parameters activate per token.
Mixture-of-Experts models contain many "expert" sub-networks but route each token through only a small subset, so the active parameter count per inference is much smaller than the total parameter count. This makes very large models tractable to serve. Most frontier 2026 models use MoE under the hood.
Related terms
Building with MoE (Mixture of Experts)?
We ship production AI systems built around concepts like this every quarter. Send a brief and get a written proposal in 48 hours.
Send a brief →