1 comment

[ 2.7 ms ] story [ 11.1 ms ] thread
The arXiv paper was submitted April 2025, the research itself isn't new, but the new is Google's blog post packaging it for a wider audience.

worth reading the original paper alongside the blog post. I think the ppaper has details the blog post glosses over, particularly around the calibration-free quantization approach and how they handle outlier channels.

Interestingly: the research sits on arXiv for a year, nobody talks about it