A model design where not all parameters are used for every computation, reducing memory and computational requirements compared to dense models.