Making a model bigger usually makes every request more expensive, because the whole thing runs each time. A mixture-of-experts design splits the model into many specialised parts and, for each input, activates only a few, so capacity grows without the cost rising in step.
A small router decides which parts handle a given input. The challenge is keeping that routing balanced so some parts are not overused while others sit idle, but when it works the approach buys scale more efficiently.