arxiv.org
> 1.58-bit FLUX achieves a 7.7× reduction in model storage and more than a 5.1× reduction in inference memory usage
Reply on news.ycombinator.com