An 8 by 8 weight matrix shown at a chosen numeric precision. A bit-depth selector switches between 32, 16, 8 and 4 bits per number. As the precision drops, a file-size bar shrinks dramatically while an accuracy bar barely dips, and the matrix cell values snap to fewer and fewer distinct levels, making the colors visibly banded. This is how giant models are squeezed onto small hardware with almost no loss of quality.

Quantization: round the numbers

Store each weight with fewer bits — far smaller, almost as smart.

storing each number with 256 possible levels

Quarter the bits, quarter the file — and the numbers still round to nearly the same values.