A low-precision numerical format that represents model weights and activations using only 4 bits, significantly reducing memory usage and inference speed at the cost of some accuracy.