A ultra-low precision format using 4-bit floating-point numbers to represent model weights, enabling extreme compression.