AI Foundationspredict · compress · act
act XI

The Open-Weight World

A file you can download is not a permission to do anything with it. The licence decides what is allowed — and the difference between licences is large.

49

Licences and Rules

frontieras of 2026-09
before this →The Compute Squeeze

When you download a model, you get two things: the weights, and a licence — the legal terms attached to them. People read the first and skip the second, then discover at launch that the terms forbid exactly what they built. This short chapter is the one to come back to before shipping anything on open weights.

read the licence on the model cardcommercial use allowed?user-count or revenue cap?redistribution & derivatives allowed?if no → stop, choose another modelif over the cap → contact the labif no → you cannot ship a derivativeall clear → track the terms anywayterms change between versions — pin the one you used
the licence decision, in order. Stop at the first box that fails.

What a licence controls

Term The question it answers
Commercial use May I sell a product built on this?
Redistribution May I host it, or ship the weights to users?
Derivatives May I fine-tune and release my version? What counts as “derivative”?
Cap Is there a user or revenue limit before I need a separate deal?
Naming Must I credit the lab, or say “built with” something?
Acceptable use What must I not do — weaponisation, surveillance, disinformation?
Jurisdiction Whose law governs the terms, and can they change them?

The common licences

Permissive (Apache 2.0, MIT) lets you use, modify, and sell almost freely, keeping the notice. DeepSeek’s R1 and Mistral’s flagship models use these — they are the reason those weights spread so fast. Community licences (Meta’s Llama, Google’s Gemma) are open in practice but carry conditions: a very large user cap, a naming requirement, and an acceptable use policy that you must pass downstream. Research-only licences forbid commercial use entirely. A model card is the document that states all of this, alongside intended use, evaluation results, and known limitations — its licence section is the part that governs you.

Open weights are not open source

the definition fight

For software, “open source” has a firm meaning, and the industry’s body, the Open Source Initiative, tried to extend it to AI in 2024. The hard question is the same as in the first chapter: if you cannot see the data or the training code, is the system open, or merely downloadable? The debate is unresolved, which is exactly why a permissive software-style licence tells you more than the word “open” does.

regulation is no longer arrivingThe EU AI Act’s duties on general-purpose AI (GPAI) providers — documentation, a copyright policy, and a summary of training content — entered into application on 2 August 2025, with further obligations for models judged to carry systemic risk. The Commission published guidelines on the scope of those duties, and a voluntary GPAI Code of Practice gives providers a route to demonstrate compliance. Other jurisdictions are adding rules of their own. The direction is uniform: the paperwork around a model now matters as much as the weights.
WEIGHTSa technical artifact
LICENCEa legal permission, attached to it
REGULATIONa public duty, above both
introduces →open-weight licenceacceptable use policymodel cardEU AI ActGPAI
← previousThe Compute Squeezenext →Running One Yourself