Hacker News new | past | comments | ask | show | jobs | submit
petu | on: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
3090/4090 probably would do 40 t/s, for 5090 75 t/s is shown in the blog.