Show HN: running GLM 5.2 on a slow laptop by streaming experts from disk

Category: Pulse
Status: discussion
Sources: Colibri (GitHub), HN discussion
Summary: A Show HN (838 points) presents Colibri, a roughly 1,300-line single-file C inference engine that runs the 744B-parameter GLM 5.2 Mixture-of-Experts model on a 12-core laptop with 25 GB RAM by keeping only the dense layers (about 17B parameters, int4, roughly 9.9 GB) resident and streaming the routed experts from disk through an LRU cache and the OS page cache. No BLAS, Python, or GPU is required.
Comments: The author reports about 0.05 to 0.1 tokens per second. Commenters debated SSD wear and whether the throughput is usable for overnight batch tasks, and compared the approach to disk-backed inference experiments.
Why it matters: It is a concrete demonstration of how far MoE routing plus paging can push large-model inference onto commodity hardware, at the cost of throughput.