How a Spare Graphics Card Adds Memory to Nvidia's Home AI Computer

The DGX Spark is Nvidia's small, quiet desktop computer built for running AI models at home instead of renting them from the cloud. It costs a few thousand dollars and comes with a fixed amount of memory — and memory is the bottleneck. The more memory you have, the bigger and smarter the AI you can run, and the more text it can hold in mind at once.

A developer named Ciprian Veg published a free tool on GitHub that solves a slice of that problem. Modern AI systems often use a clever shortcut called speculative decoding: a small, fast AI drafts an answer, and the big AI checks and approves it. That little draft AI eats up memory on the Spark. This tool moves it onto a spare graphics card you might already have, anywhere from 10 to 24 gigabytes. The freed memory goes back to the main AI, letting it read longer documents or run a less compressed, smarter version of itself.

The catch is obvious. This only helps if you already own a DGX Spark and a spare graphics card, which most people don't. It's a tool for hobbyists and small teams tinkering with local AI, and setting it up takes some comfort with the command line. There's no app store button here.

Still, the direction matters. Running AI on your own desk means no monthly bills, no sending your data to a company's servers, and no internet required. Each small improvement like this makes that option a little more practical — and nudges the whole field closer to the day when owning your own AI assistant is normal, not exotic.