A 744B AI Model on a Laptop With No GPU

A small engine with a big appetite

The most capable AI models are usually far too big for a normal computer. They live in data centres, and you rent access to them.

Colibri, made by Vincenzo Fornaro, takes a different route. Its promise is simple: run frontier models "on hardware you already own." His tagline is lovely too: "Tiny engine, immense model."

How it pulls this off

Many big models today are built as a team of specialists. When you ask a question, only a few of those specialists actually answer. The rest sit quietly.

So Colibri doesn't keep the whole team in the computer's short-term memory. It leaves them on the hard drive and calls in the ones it needs, when it needs them. Think of a kitchen: you don't keep every ingredient on the counter, you grab them from the walk-in fridge as each dish comes up.

It's also built with no outside parts to install, which keeps it light.

People noticed. The project roughly doubled its GitHub stars between two snapshots, landing at 41.6k. Fornaro has since spun off lumabri, which spreads big models across several of your own machines.

Words worth knowing

Mixture of experts (MoE): an AI model made of many specialists, where only a few work on each request.

Weights: the learned knowledge inside a model, stored as a very large file.

Running locally: the AI works on your own machine, so your data stays with you.

Something to sit with

If a serious AI could live in your office, never sending a client file anywhere, what would you finally let it help with? Have a peek: github.com/JustVugg/colibri