16 August 2026

gemma4 12b

My PC is starting to show its age. I bought it after selling my motorcycle, back when I finally gave up on the idea of making a full physical recovery from POEMS Syndrome secondary to Castleman’s Disease, which I had been diagnosed and treated for.

It wasn’t exactly the purchase I originally envisioned, but this computer has served me pretty well over the years.

It’s a custom-built system with a 3rd-generation AMD platform, an MSI motherboard, 32 GB of RAM, and a Ryzen 7 3700X processor. Originally, I was running an 8 GB RTX 2060, but in late 2025 I picked up a used Radeon RX 6750 XT from eBay. That GPU gave the system a pretty significant upgrade and bought me some additional life out of what is otherwise an aging machine.

Recently, I’ve been experimenting with running AI models locally, and that brought me to Google’s Gemma models. Looking through the available Gemma downloads, I was curious about how far I could push my current hardware.

Despite the age of my system, the specifications looked like they should be enough to handle the 12B model reasonably well. Of course, “should handle it” and actually using a model comfortably are two different things, so I decided to give it a try rather than simply assuming my hardware wasn’t up to the task.

I also decided to include a brief video demonstrating its response. Rather than just telling you that it works, I figured it would be more interesting to actually show the model running and responding on my aging hardware.

In my opinion, this is one of the more interesting things about running AI locally. You don’t necessarily need a brand-new, top-of-the-line computer to experiment with these models. Older hardware can still have some life left in it, especially if you’ve upgraded individual components along the way.

My PC certainly isn’t what I would consider a modern AI workstation, but if it can run a 12B model well enough to be useful, that’s pretty impressive considering how old the rest of the system is becoming.

I’m still experimenting with local AI and figuring out what these models can—and can’t—do on hardware that most people would probably consider outdated. But that’s part of the fun.

Sometimes you don’t need the newest hardware. You just need to see what you can get out of what you already have. Peace and bacon grease.