Colibri

Colibri is a lightweight engine that runs GLM-5.2, a 744-billion-parameter model, on a consumer machine with 25GB of RAM. It gets there by streaming expert models straight off disk instead of holding everything in memory, which is what keeps memory use and performance workable on ordinary hardware.

The engine is written in pure C and does not drag in a complex dependency chain. That combination is aimed at AI enthusiasts, developers, and researchers who want to work with a large model without buying high-end hardware first.