Using fewer bits to represent numbers in a model, which speeds up computation and reduces memory usage at the cost of some accuracy.