Developer builds Colibrì to run GLM 5.2’s 744B MoE model on a 32GB RAM computer without a GPU

A developer says they were able to run GLM 5.2 locally on a modest 12-core machine with 32GB of RAM by converting the model to int4 and streaming its routed MoE experts from disk. Their project, Colibrì, keeps the dense portion in RAM while loading expert weights on demand, enabling inference without Python, BLAS, or a GPU, albeit at very slow speeds.

The project highlights how Mixture-of-Experts architectures can be exploited for extreme local inference by avoiding full-model residency in memory. Even if performance is impractical today, it points to growing interest in lightweight runtimes, disk-backed inference, and experimentation that broadens access to frontier-scale models beyond high-end GPU systems.