A 4-bit floating-point number format that represents model weights with very low precision, enabling extremely efficient inference on compatible hardware.