Along with the ParetoQ paper from Meta (
https://arxiv.org/abs/2502.02631), the concept of low-bit LLMs seems to be gaining traction. Has anyone experimented with this in production? I'm aware of a few pre-transformer era companies focused on applying this to CNNs
