Quantization isn’t a shortcut. It’s a test.

More training leads to better compression. Trimming precision off the integers reveals that there was more compression that could be had.