Linux How To: I Ran the Huge GLM 5.2 Model on a “Too Small” Linux Laptop Using Colibri

Reference links: https://github.com/tim-dickey/my-playbook-runbook-and-other-artifacts-repo/blob/main/runbooks/colibri-pangolin15-runbook.pdf https://youtu.be/FwmNucVMrfs?si=lIaGVmH2JCnlrmiA https://github.com/JustVugg/colibri https://huggingface.co/mateogrgic/GLM-5.2-colibri-int4-with-int8-mtp In this how-to video, I walk through how I used Colibri to run the massive GLM 5.2 model on a Linux notebook that should not, on paper, be able to handle it—and why that matters if you’re a home lab tinkerer or aspiring AI builder. I start by introducing the GLM model and explaining that it’s well beyond what my local hardware is officially designed for, then give a big shout-out to Vincenzo, whose work on Colibri makes this experiment possible. You see me remote into my @System76 Pangolin 15 Linux laptop using RustDesk, the same machine where I previously “hacked” the integrated GPU to appear as though it had 32 GB of VRAM by unlocking up to 96 GB of system RAM for shared graphics memory. I reference that earlier video and explain how this hardware hack set the stage for trying something as ambitious as running GLM 5.2 locally. From there, I show the reality behind the scenes: readme files that assume deeper expertise, GitHub runbooks I’m creating and publishing so you can follow along, and plenty of false starts—including failing to create directories inside a virtual environment and needing to drop back to a root terminal to set things up correctly before starting the download. I stress that what you’re seeing in the recording is the back half of at least thirty minutes of pre-work and troubleshooting, much of it assisted by Perplexity, which I use as my primary research copilot to decode the multi-step instructions and get to a working install command. I explain that downloading GLM 5.2 via Colibri is not quick: in my case, it took overnight to complete, even on a fiber connection with the laptop on Wi‑Fi, and I warn you to expect a similar “set it and walk away” experience. Once the download finishes, I briefly show how I validate that the model is really there by inspecting file sizes and hashes, and mention that I’ll include the runbook link in a GitHub repo so you can see the exact commands and validation steps. The real payoff comes when I actually start the GLM 5.2 model in its Colibri container and send it a simple prompt: “Tell me a short story about a Viking explorer.” We watch in real time as the model begins to respond, slowly, while I pull up system resource monitors to show what’s happening under the hood: memory pegged, CPU heavily utilized, and the console process consuming the bulk of available resources because it’s running the model. I note that the mobile GPU isn’t really participating here; this is largely CPU-bound, which is why the fan noise and sluggishness are so pronounced. Throughout, I emphasize that this is a “hacky” experiment that works, but comes with plenty of rough edges and moments of frustration—especially if, like me, you’re a technologist and product person rather than a professional Linux engineer or ML infra expert. I’m careful not to oversell the ease: there are janky parts, the documentation is often written for developers, and you’ll likely run into your own walls. The point is not that it’s “easy,” but that it is possible, and that modern tools (Colibri, open models, Perplexity, good hardware) let an individual push far beyond what their laptop’s spec sheet would suggest. I wrap up by encouraging you to try this yourself if you’re curious and ready to learn, framing it as a great personal investment in understanding how large language models actually run—especially on hardware that wasn’t originally built for this class of workload. I also point you to my Olares Project playlist for more episodes in the same spirit: experiments, wiring, and honest looks at what it takes to make advanced AI systems work in a home lab context.