Open-Weight Model

An AI model whose weights, the "numerical parameters within an AI model that helps determine the model's output in response to inputs" in the wording used by the US National Telecommunications and Information Administration (NTIA), have been "released openly to the public, either by allowing users to download them from the Internet or through other mechanisms". The contrast is a closed model offered through a web interface or an API, where users "can interact with these models by submitting inputs and receiving outputs, but cannot directly access the weights themselves". Once weights are out, NTIA notes, "individuals and firms can customize them outside the developer's initial scope", for example by fine-tuning on new data, which typically takes far less expertise and computing power than training a model from scratch. The openness cuts both ways: "Adversarial actors can remove safeguards from open models via fine-tuning". Release terms vary too, since "some licenses require the user or downloader to agree to use and redistribution restrictions". For businesses, an open-weight model can run on infrastructure the company controls, but checking the licence, and owning the safeguards and security, then falls to the company rather than the supplier.