It’s local. You can tell by how your computer dies from inference, when you upload new assets 😅. And you can run the models on another computer separately and watch that one die then, when you upload new assets.