In a mixture-of-experts model, an expert is one of several parallel sub-networks (typically feed-forward layers) that each specialize in different types of input. A router selects which experts activate for each token, allowing the model to scale total parameters without proportionally scaling compute per token.