Model weights are the numbers inside a trained AI model that hold what it learned. When people download or share a model, the file is mostly these weights. Think of a bakery: the oven is the software, but the recipes are what make the bread taste right. Whoever has the weights can run the model on their own computer or server. When a maker publishes the weights openly, the result is called an open-weight model.
Why it matters
Weights matter for safety in two ways. First, a model file from an untrusted source can be dangerous. Some file formats can carry malicious code, and a model can also hide a backdoor that makes it misbehave in ways you would not notice. Both are forms of an AI supply chain attack. If a staff member or volunteer wants to download a model, ask where it came from, whether the source is trustworthy, and whether its file format is one that can carry code. Second, if a vendor runs a model for you, ask who holds the weights and who else could get a copy, since anyone with the weights can run the model.
Published Updated