744B on a Laptop? colibrì Actually Runs It (25GB RAM)

colibrì runs a 744-billion-parameter GLM-5.2 frontier model on just 25 GB of RAM — no GPU, no cloud, pure C streaming from disk. In this deep dive, we break down how the tiny C engine makes the impossible possible. Mixture-of-experts architecture means only ~40B parameters fire per token. The resident dense core (~9.9 GB) stays in RAM while the remaining 370 GB of routed experts stream off your SSD on demand. Multi-head latent attention compresses the KV cache by 57×. Speculative decoding with an int8 MTP head pushes useful latency down. And the engine literally learns you — recording your expert usage to pin your hottest paths into RAM. We cover the architecture, the disk-to-RAM pipeline, MLA compressed KV, MTP speculative decoding, the learning cache, real community benchmarks, and the honest truth about what colibrì is: a 744B model answering where nothing else can. Slowly, but it runs. ⏱ CHAPTERS 0:00 — Intro 0:04 — The Impossible 0:20 — Why Clusters 1:31 — Architecture Deep Dive 3:50 — It Learns You 4:58 — How To Run 7:00 — The Honest Take *Repo:* https://github.com/JustVugg/colibri *Pre-converted int4 model:* https://huggingface.co/jlnsrk/GLM-5.2... *Int8 MTP heads:* https://huggingface.co/mateogrgic/GLM... colibrì v1.0 — Apache 2.0 engine, MIT weights. Pure C, no BLAS, no Python at runtime, no GPU required. Compile with gcc and OpenMP, set COLI_MODEL, run `coli chat`. Under a minute to a prompt. If you run it, post your benchmark numbers to the repo — the community is building the real-world speed database as we speak. #OpenSource #AI #colibri #GLM5 #LocalLLM #EdgeInference #MachineLearning #DeveloperTools #opensource #aiterminal #LLM