Is there any case of a less quantised smaller model outperforming a more quantised larger model?(reddit.com)