Tools · Hacker News ·
Show HN: Getting GLM 5.2 running on my slow computer
A developer describes getting GLM 5.2 running on a 32GB RAM computer by converting the model to int4 and using disk-backed expert loading. The setup, called Colibrì, targets a 744B MoE model that activates about 40B parameters per token and runs slowly but fits without out-of-memory errors.