GLM-5.2 With No GPU: We Tested Colibrì's 0.1 tok/s Claim

Colibrì is a pure-C engine that runs GLM-5.2 — Z.ai's 744-billion-parameter mixture-of-experts model in about 25 GB of RAM by streaming the ~370 GB of expert weights off a local disk through the OS page cache. The author's own number is 0.1 tokens per second, cold. We didn't repeat that number we tested it ourselves, on our own Dell PowerEdge R730 (128 GB RAM, SATA SSD mirror, no NVMe): 0.02 tok/s cold, 0.06 tok/s warm, and a RAM-budget sweep from 8 to 44 GB that made no measurable difference. Then the control experiment: the same model through stock llama.cpp with a Tesla V100 and MoE expert offload — cold, it spent 40+ minutes on one sentence. The wall is the disk, not the engine. Chapters, sources, and our full benchmark table below. AI transparency: narration and anchor visuals are AI-generated (our narrator Wolf Stedman is synthetic). Every benchmark number is a real measurement on our own hardware; terminal footage is real screen capture, time-lapsed where labeled, with the original elapsed clock burned in. Sources: Colibrì repo: https://github.com/JustVugg/colibri Show HN thread: https://news.ycombinator.com/item?id=... Weights (int4, ~370 GB): https://huggingface.co/jlnsrk/GLM-5.2... Gigazine coverage: https://gigazine.net/gsc_news/en/2026... Unsloth GLM-5.2 guide (GGUF route): https://unsloth.ai/docs/models/glm-5.2 0:00 Colibrì's claim: GLM-5.2 (744B) in 25GB of RAM 0:23 The old hardware floor: 256GB 1:15 How Colibrì works: disk-streamed MoE experts 2:23 The Show HN moment 3:08 What 0.1 tokens per second actually means 4:26 The llama.cpp + GGUF route and what it needs 5:42 Our benchmark: cold vs warm + RAM sweep 7:41 Control test: llama.cpp + Tesla V100, same disk 9:08 Verdict: the wall is the disk, not the engine 10:32 What we test next