The Colibri Playbook — Run GLM-5.2, a 744-Billion-Parameter Model, on the Machine You Already Own

The companion to the full video. Colibri is a pure-C, zero-dependency engine that runs GLM-5.2 (744B Mixture-of-Experts) on a machine with ~25 GB of RAM by treating VRAM, RAM and SSD as one memory hierarchy: the dense core (~17B params at int4, 9.9 GB) stays resident, and all 19,456 experts (~19 MB each, ~370 GB) stream from disk the moment the router asks — with a learning cache that pins YOUR hottest experts so it gets measurably faster the more you use it. THIS PLAYBOOK: the three-shelf placement model on one page, the exact setup for Linux, Windows and Mac (including the check-your-machine-first commands coli doctor and coli plan), the int8 draft-head trap and the exact file sizes to verify before you download 370 GB, the honest speed ladder from the 25 GB floor (0.05–0.1 tok/s) to a 128 GB CPU box (~1.8 tok/s warm) to full GPU residency (5.8–6.8 tok/s), and all 8 tuning knobs that matter — PIPE, PILOT, PIN_GB, CUDA_EXPERT_GB, SPEC_PIN, DRAFT, GRAMMAR, COLI_NUMA — explained in plain English. Every number verified against the colibri README on July 21, 2026. Engine by JustVugg (Apache 2.0); GLM-5.2 weights by Z.ai (MIT). Independent field manual from Hyperautomation Labs.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.