Open weight AI models are the ones you can download and run yourself, because the company published their weights. Which raises a fair question: what's a weight?
Picture a mixing board. Training a model means pushing enormous amounts of data through it and nudging every knob a hair each time the output lands closer to right. When training stops, the knobs freeze. Those frozen positions are the weights, and they're the whole model.
The board below has six knobs instead of a few hundred billion, but it works the same way. The dashed line is a pattern the model is trying to match. The solid line is what it currently predicts. Press Train, then drag a knob and watch yourself break it.
Six numbers. A frontier model is the same idea with hundreds of billions of them, which is why the download runs to hundreds of gigabytes. Publishing those numbers is what open weights means.
Because a file behaves differently than a service. If you have the weights, you can run the model on your own hardware, adapt it to your own work, keep your data in house, and switch providers without starting over. And once those numbers are published, nobody can call them back, which is the tradeoff at the center of the open weights declaration signed by twenty-five organizations in July 2026.