Colibri: Running GLM-5.2 on a 25 GB Laptop and How Slow It Really Is
By Matt Cullerton, founder of Mavric and Apso
Colibri is an open-source C inference engine that runs Z.ai's 744-billion-parameter GLM-5.2 model on a CPU-only machine with 25 GB of RAM by streaming the model's experts from disk. A TikTok from @agilesingh says it puts Nvidia's "whole business model" in question. The engine and its numbers are real, but the video leaves out the speed: the project's README calls 0.05 to 0.1 tokens per second on that laptop "the honest baseline," and its fastest published results come from six Nvidia GPUs.
What is Colibri?
Colibri (stylized "colibrì," Italian for hummingbird) lives at github.com/JustVugg/colibri under the Apache-2.0 license. Its author is Vincenzo Fornaro, whose GitHub bio reads "Founder of Colibrì a tiny engine, immense model for affordable hardware." The README opens with "Pure C, one file per model family, no GPU required."
The video calls it "one guy's laptop." That was true at the start. The README says "colibri started as a one-person project on a 12-core laptop with 25 GB of RAM; today its numbers come from a community of real machines." When the video was filmed the repo had 2,230 commits, 3.5k forks and 17 releases, including merged pull requests from other contributors.
How does Colibri run a 744B model on 25 GB of RAM?
GLM-5.2 is a mixture-of-experts model. The README explains: "A 744B Mixture-of-Experts model activates only ~40B parameters per token and only ~11 GB of those change from token to token (the routed experts)." Forty of 744 billion is about 5.4 percent of the model active per token.
Colibri splits the model in two:
- The dense core (attention, shared experts, embeddings, about 17 billion parameters) stays in RAM at int4, which takes about 9.9 GB.
- The 19,456 routed experts, about 19 MB each at int4, stay on disk and are streamed on demand, with a per-layer cache and an optional GPU tier.
A thread called PILOT prefetches the next layer's experts. The README says "GLM-5.2's routing is 71.6% predictable one layer ahead."
Where does the 19,456 figure come from?
It is exact. The README derives it as "75 MoE layers x 256 + the MTP head." GLM-5.2's config.json on Hugging Face lists 78 layers, 256 routed experts and first_k_dense_replace: 3. Seventy-eight minus three dense layers leaves 75 expert layers, and 75 times 256 is 19,200. The 256 experts of the multi-token-prediction head bring it to 19,456.
The config also lists num_experts_per_tok: 8, meaning each token uses 8 of 256 experts per layer. That small number is why so little data changes per token and why streaming from disk is workable at all. The video never mentions it.
How fast is Colibri without a GPU?
Very slow on the 25 GB machine. The README at v1.11.0, the version on screen, says: "25 GB dev box: 0.05 to 0.1 tok/s cold, the proven floor where this project started, and still the honest baseline." At that rate one token takes 10 to 20 seconds, and a 500-token answer takes 1.4 to 2.8 hours. A GIGAZINE writeup from 2026-07-10 reports the same range.
The README treats the GPU as a speed setting: "A GPU only ever makes it faster. Speed is set by your disk." Its published GLM-5.2 figures at v1.11.0:
- 6x RTX 5090, full residency: 5.8 to 6.8 tok/s, and 9.0 to 9.2 with selective NUMA interleave in the current README.
- 128 GB CPU-only desktop: about 1.8 tok/s warm.
- Single RTX 5070 Ti box: 1.07 tok/s.
- 25 GB laptop: 0.05 to 0.1 tok/s cold.
The terminal footage in the video shows a different model. The banner reads curie-17b-alpha, a separate 17B model the same author trained, running at 49.5 and 84.9 tok/s with visibly broken output. No frame shows GLM-5.2 generating text.
How big is the GLM-5.2 download?
The video says 372 GB, which matched the README at the time. The weights are a community int4 conversion by Mohamed Mastouri on Hugging Face, not an official Z.ai file. That page now lists about 429 GB: 390.5 GB of expert weights across 141 shards plus a 9.3 GB MTP head. The current README also says 429 GB.
At 100 Mbit/s, 429 GB takes about 9.5 hours to download. Running the model is free of API charges, but it costs disk space, 22 to 25 GB of RAM held for the session, and hours of sustained NVMe reads.
Does Colibri threaten Nvidia?
Nvidia's own numbers say no. Its press release of 2026-08-26, for the quarter ended 2026-07-26, reports revenue of $96,221 million and data center revenue of $89.0 billion, up 117 percent year over year, with 75.0 percent gross margins and $108.0 billion of guidance for the next quarter. That filing came one month before the video.
Colibri's own speed ladder is built from Nvidia cards. Six RTX 5090s at about 9.1 tok/s are roughly 121 times faster than the laptop at 0.075 tok/s.
The order of events also runs against "just got questioned." Z.ai published GLM-5.2 under an MIT license, with its model card blog dated 2026-06-17. Colibri had run in production since late June and tagged v1.0.0 on 19 July. Nvidia reported in August. The video posted on 2026-09-26, about three months after Colibri appeared. The repo has since reached v2.0.0 with 42.9k stars and engines for GLM-5.3, Kimi K3, DeepSeek V4 and others.
Bottom line
Colibri does what it claims: it runs a 744B model on a CPU-only laptop by streaming experts from disk, and its 19,456 figure is exact. On that laptop it produces one token every 10 to 20 seconds. Plan for 429 GB of disk, and add a GPU if you want answers in minutes.
Sources
- JustVugg/colibri on GitHub - current repo, benchmark ladder and 429 GB figure.
- Colibri README at v1.11.0 - the version on screen, with the expert derivation and speed floor.
- Vincenzo Fornaro's GitHub profile - author name and bio.
- Colibri releases - the v1.0.0 date and production note.
- Colibri quick start - RAM requirements and the warning that generation can be under 1 token per second.
- GLM-5.2 int4 weights on Hugging Face - the community conversion and its size.
- Z.ai config.json for GLM-5.2 - layer and expert counts.
- GLM-5.2 model card - license and launch date.
- Nvidia Q2 FY2027 results - revenue, data center growth and guidance.
- GIGAZINE on Colibri - independent report of the 0.05 to 0.1 tok/s rate.
Published September 26, 2026. Based on a TikTok by @agilesingh. Full video notes (markdown)