Live23 Sept 2026/8 verified items today/Claim check: Yes, a 744B model runs on a laptop with no GPU/Daily digest
aimluptodate
My Stack
Preview: items marked "(demo)" are sample data, not yet editor-verified. How we verify

← Claim Check

The claim · 23 Sept 2026

A developer ran a 744B-parameter AI (GLM 5.2) on a laptop without a GPU.

Seen on: YouTube Short: "Calibri runs 744B parameter AI on laptop without a GPU" ↗

True, with a catch

Yes, a 744B model runs on a laptop with no GPU. No, you won't want to chat with it.

Colibrì really does run GLM-5.2 (744B total, ~40B active per token) with 16–25 GB of RAM and no GPU. The catch: it needs ~372 GB of fast SSD, and on a 25 GB machine it produces roughly one token every 10–20 seconds.

Disagree / report an error

The numbers

Total parameters744B~40B active per token
Vendor-documented
RAM16–25 GB24 GB comfortable
Vendor-documented
Disk~372 GBfast NVMe recommended
Vendor-documented
Speed, 25 GB box0.05–0.1 tok/s1 token every 10–20 s
Community-reported
Speed, 128 GB desktop~1.8 tok/sCPU only, warm cache
Community-reported
Cost per token$0after hardware and electricity
Vendor-documented

What's true

  • The engine is Colibrì: a small, open-source (Apache-2.0) C engine with no Python at runtime and no GPU required. Vendor-documented
  • GLM-5.2 is a Mixture-of-Experts model from Z.ai: 744B parameters in total, about 40B active for each token. Vendor-documented
  • Only the shared part of the model (~17B params, int4, ~9.9 GB) stays in RAM. The 19,456 routed experts (~370 GB) live on disk and are streamed in on demand, with a per-layer cache. Vendor-documented
  • It works fully offline and can serve an OpenAI-compatible API on your own machine, so prompts never leave it and there is no per-token bill. Vendor-documented

The catch

  • Speed. On a 25 GB-RAM machine with a cold cache: ~0.05–0.1 tokens/sec, i.e. one token every 10–20 seconds. A 300-word answer can take over an hour. Community-reported
  • Disk. You need ~372 GB free, ideally on a fast NVMe SSD; disk bandwidth sets the speed. Vendor-documented
  • More hardware helps a lot: ~1.8 tok/s on a 128 GB CPU-only desktop, ~1 tok/s with a single laptop-class RTX 5070 Ti, 5.8–6.8 tok/s on 6× RTX 5090. Community-reported
  • It is a proof of capability, not a production setup. The project itself offers no speed guarantees. Vendor-documented

What the post got wrong

  • The project is called Colibrì (Italian for hummingbird), not "Calibri".
  • "Without a GPU" is true; "on a laptop" is true only for laptops with enough free NVMe space, and "usable" depends heavily on your patience.

How it works, in plain words

Mixture-of-Experts models only use a small slice of their weights for each token. Colibrì keeps the always-used part in RAM and fetches just the experts each token needs from disk, like a library that keeps the popular books on the desk and fetches the rest from the stacks. That trades speed for memory: disk reads replace GPU memory.

Worth trying if you are…

  • Researchers and tinkerers who want frontier-scale weights fully offline
  • Batch jobs that can run overnight: summarising a document set, one closed question at a time
  • Anyone evaluating whether a big open-weight model beats their small local model on quality

Skip it if you need…

  • Anyone who needs interactive chat or coding help. A 7B–32B model in Ollama or LM Studio will feel instant by comparison
  • Production workloads with latency targets
  • Machines with a slow SATA SSD or a nearly full disk

Try it (Linux, macOS or Windows 11)

  1. Check you have ~372 GB free on an NVMe SSD and at least 16 GB of RAM (24 GB is comfortable).
  2. Build the engine from source (or download a prebuilt release).
  3. Download the int4 GLM-5.2 weights for Colibrì from Hugging Face (372 GB).
  4. Start the chat UI, or `serve` for an OpenAI-compatible local API.
git clone https://github.com/JustVugg/colibri && cd colibri/c
./setup.sh

# weights: huggingface.co/mastouri/GLM-5.2-colibri-int4-g64-with-int8-mtp (372 GB)
COLI_MODEL=/nvme/glm52_i4 ./coli chat     # chat in the terminal
COLI_MODEL=/nvme/glm52_i4 ./coli serve    # local OpenAI-compatible API

Sources

Our own test

Scheduled: we'll measure tokens/sec on a 32 GB laptop (NVMe) and a 128 GB desktop, plus quality vs. a 32B local model on our RAG suite.

See it in the Daily Pulse
Want something that works today?

Run powerful AI on my own laptop

From zero to a private local assistant in about 15 minutes, and a clear picture of how big you can go.

Open the guide