This model has fast weights, which actually are modified during inference.