← The AI Hype Audit — all 325 verdicts

PARTLY Partly true: Colibri really does run the 744B GLM-5.2 on a 25 GB-RAM machine with no GPU, but at about one word every 10 to 20 seconds off a ~372 GB download, not as a minutes-to-set-up replacement for your OpenAI bill Facebook · Sep 2026

The claim'Nvidia's empire is sweating right now.' A developer's open-source project, Colibri, runs GLM-5.2, 'a huge 744 billion parameter model that beats paid AIs', entirely offline on a standard laptop with 25 GB of RAM and no graphics card, by keeping the active part in RAM and streaming its experts from the SSD. 'Setup takes minutes', 'you have a world-class AI running locally', and pointing your code's base URL at it means 'dropping your OpenAI bill to zero.' 'Comment AI and I'll send you the GitHub link and full install guide.' The reel's own caption adds: the model container needs roughly 400 GB of storage and the original laptop managed only 0.05 to 0.1 tokens per second when cold, 'a proof of access, not instant laptop inference.'

The engineering is real and genuinely clever. Colibri (github.com/JustVugg/colibri, Apache 2.0, about 37.5K stars when checked) is a pure-C engine that treats VRAM, RAM and NVMe as one memory ladder: GLM-5.2's shared weights (about 9.9 GB at int4) stay in RAM and the routed experts (about 370 GB) are pulled from disk as each token needs them. GLM-5.2 is real too: Zhipu/Z.ai's open-weight MoE, about 744B total and about 40B active per token, MIT license, released in June 2026 and strong on coding benchmarks. What the voiceover does is strip out the numbers that the reel's own caption and the repo publish. The 25 GB box managed 0.05 to 0.1 tokens per second cold, so one sentence can take several minutes. The repo's own table has a 128 GB desktop at about 1.8 tok/s and six RTX 5090s at about 6 tok/s. 'Setup takes minutes' ignores a ~372 GB model download and a fast NVMe drive to hold it. The OpenAI-compatible server is real, but at those speeds it cannot stand in for a paid API on any real workload, so your OpenAI bill does not drop to zero. And Nvidia is not sweating over a tool whose fastest published setup runs on six Nvidia cards. 'Comment AI for the link' is a DM funnel for a public GitHub repo.

What holds up

  • Reel read frame by frame: on-screen captions ('Nvidia's empire is sweating', 'took GLM 5.2', 'beats paid AIs'), the Colibri web dashboard, a benchmarks.md page, and the repo card 'JustVugg/colibri' showing about 21K stars and 2K forks at recording time.
  • github.com/JustVugg/colibri (fetched Sep 24, 2026): pure C, zero runtime dependencies, Apache 2.0, about 37.5K stars and 4.1K forks. GLM-5.2 container about 372 GB at int4, 16 GB RAM minimum and 24 GB comfortable. Measured speeds: 25 GB dev box cold 0.05 to 0.1 tok/s, 128 GB CPU desktop warm about 1.8 tok/s, single RTX 5070 Ti 1.07 tok/s, 6x RTX 5090 5.8 to 6.8 tok/s. Ships an OpenAI-compatible `coli serve`.
  • GLM-5.2 verified in multiple independent write-ups (DataNorth, NxCode, MindStudio and others): Zhipu AI / Z.ai, about 744B total and about 40B active parameters, 1M context, open weights on Hugging Face under MIT, launched June 13, 2026. Its wins over GPT-5.5 are on specific coding and long-horizon benchmarks, not across the board.
  • The '21,000 experts' in the voiceover is the caption's 19,456 routed experts (76 layers x 256), rounded up. Several forks with the same description exist on GitHub, so a DM'd link is not guaranteed to be the original.

What doesn’t

  • The voiceover drops the two numbers that decide everything, ~372-400 GB of storage and 0.05-0.1 tokens/sec cold, which the reel's own caption admits.
  • 'Setup takes minutes' and 'drop your OpenAI bill to zero' are false in practice: the download alone takes hours on most connections, and a local model this slow cannot serve a real app or team.
  • 'Nvidia's empire is sweating' is rage-bait; the project's fastest published configuration is six Nvidia RTX 5090s.
  • Comment-for-the-link funnel for a public GitHub repo; forks with the same description exist, so go to the original yourself.

The catch

It proves a 744B model can run on hardware you already own. It does not make that model usable. At 0.05 to 0.1 tokens per second cold, a one-paragraph answer can take a quarter of an hour, and you first need about 372 GB of free NVMe and a very long download. It is a great experiment and a bad plan for replacing a paid API.

How to actually do it

  • If you want to try it, go straight to github.com/JustVugg/colibri (not a DM'd fork), confirm you have ~400 GB free on a fast NVMe SSD, and read the repo's benchmarks page before you start the download.
  • If what you actually want is a lower AI bill, run a smaller open model that fits in your RAM or GPU, or call GLM-5.2 through Z.ai's or a hosting provider's paid API, which reporting puts at a fraction of frontier-model prices.
  • Before switching any real workflow, time a real task: point your code's base URL at the local server, run 10 typical requests, and compare speed and quality with what you pay for now. Keep the paid key as a fallback.

Colibri and GLM-5.2 are both real, and the no-GPU, 25 GB-RAM run is real. The voiceover's 'minutes to set up, world-class AI locally, OpenAI bill to zero' pitch skips the ~372 GB download and a cold speed of a few tokens per minute, which the reel's own caption concedes.

Confidence
High
Posted by
Automatrix, an AI-tools explainer page (~292K views on this reel); reel sent to Buddy Wed, Sep 23, 2026, ~9:36 PM

See the original claim →

We test hype for free. We build the real thing for a living.

Thirty minutes, no pitch — and you'll leave with something useful either way.

Book a call with Todd or start with the free Business Checkup →

The Verdict Weekly

Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.

© Schreier Group · schreiergroup.com · See a wild AI claim? Drop it here and we'll test it.