Compressing the model dramatically slows down model read and write
with a little benefit: model compression ratio is very low.