They're using GTPQ -- here you go: https://arxiv.org/abs/2210.17323 . The authors benchmarked two families of models over a wide range of numbers of params.