Mixture of experts is a model architecture that routes inputs to selected specialist subnetworks or components rather than using the entire model for every token or example.
The approach can improve training or inference efficiency at scale, but routing quality, load balancing, and deployment complexity become important design concerns.