people learn the following two terms and use them to sound smart 1. prompt caching 2. quantization breaking prompt cache is ok. you can't guess what will happen you have to look at data at scale from real users the highest quality providers for a couple of models serve it at FP4, there's way more to quality than quantization