A mechanism that selectively activates only a subset of model experts for each input, reducing computation while maintaining specialization.