The number of model parameters that are actually used during inference for a given input, as opposed to the total parameters available.