Infrastructure and sovereignty · Layer 4
Open-weight model
An AI model whose “weights”, the settings it learned, are published: you can run it yourself, within the limits of its licence.
A way to picture itIt is like receiving a ready-to-fit engine without the factory plans or the parts list: you can install and tune it, but not rebuild it. The delivery note sets out what you are allowed to do with it.
In practiceOpen does not mean free of conditions: you gain control, and you take on the running of it.
A procurement department is comparing two open-weight models to analyse supplier contracts in-house. The first is published under the Apache 2.0 licence; the second comes with an acceptable use policy and reserves certain rights to its developer. Before choosing, the legal team checks three points: commercial use, redistribution terms and excluded regions. The licence weighs as much as performance in the decision.
Under the AI Act, the developer of a model released under a free and open-source licence is spared part of the documentation duties, provided the licence allows access to the model, its use, modification and redistribution. The copyright policy and the training-content summary still apply, and there is no exemption for a model with systemic risk (EU · AI Act, art. 53). For your part, you remain responsible for how you use it: your staff’s AI literacy (EU · AI Act, art. 4) and deployer obligations if the use is high-risk (EU · AI Act, art. 26). Hosting the model yourself keeps your data in-house, but security (EU · GDPR, art. 32) and your record of processing (EU · GDPR, art. 30) remain your job.
“Under which exact licence is this model published, what does it rule out for our use, and who on your side tracks its security updates?”
Not necessarily. The open source AI definition published in October 2024 (United States · Open Source Initiative) requires the weights, the complete code and enough information on the training data to build an equivalent system. Most so-called open models publish only the weights.
Yes. The Llama 3.2 acceptable use policy (United States · Meta) does not grant rights over its multimodal models to companies whose principal place of business is in the European Union, except to end users of a product that includes them. Hence the need to read each licence and its use policy, version by version.
Rarely for a simple adjustment. Guidelines from the European Commission (EU · European Commission, July 2025) presume a significant modification when it uses more than a third of the compute used to train the original model. You do, however, remain responsible for the system you build with it.
Educational explainer, not legal advice. Up to date as of 29 October 2026.