Learn
What is a mixture of experts model?
Last reviewed July 16, 2026
A mixture of experts (MoE) model divides its feed-forward layers into many parallel subnetworks (experts) and routes each token through only a few of them. Only a fraction of the model's total parameters activate per token, so an MoE gets the quality benefits of a very large parameter count while spending the compute of a much smaller model on every request.
Why parameter counts mislead
MoE model cards quote two numbers: total parameters and active parameters per token. A model listed at 1T total with 32B active computes like a 32B model at inference while drawing on the larger learned capacity. Comparing an MoE's total count against a dense model's count overstates its serving cost; active parameters predict speed and cost per token better.
The trade-offs
MoE wins throughput per dollar at scale, which is why many frontier and open-weight flagships use it. The cost: all experts must sit in memory even though few run per token, so memory needs track total parameters. That makes large MoEs efficient for providers with big GPUs and awkward for local use, where your RAM must hold the full model.
idapt's catalog lists dense and MoE models side by side with per-token prices and benchmark scores, so you compare on measured capability and cost rather than parameter counts.
Frequently asked
See it in practice
The fastest way to understand it is to use it. Start free, no credit card.
Try idapt free