What Is an AI Model License? Usage Rights and Restrictions

Model licences receive attention at procurement and rarely afterwards, which is a problem because the restrictions they contain frequently bite later — at scale, at a use case nobody anticipated, or when someone wants to train on outputs. This explainer covers what to check. It complements what is open-weight ai and ai vendor evaluation checklist, and reflects FISTA Solutions' approach in AI enablement delivery. This article is general guidance, not legal advice.

Does open weights mean open source?

No, and the conflation is widespread. Many prominent models publish downloadable weights under bespoke licences that carry restrictions an open source definition would exclude: prohibited use cases, thresholds above which different terms apply, naming obligations, and limits on what may be built with outputs.

Downloadable is a technical fact. Permissive is a legal one, and they frequently do not coincide.

Term type	Typical content	Bites when
Prohibited uses	Named categories excluded	Use case expands
Scale threshold	Different terms above a level	Product succeeds
Attribution	Naming in derivative products	Marketing decisions
Output training limits	Cannot train competitors	Distillation attempted
Pass-through	Terms bind your users	Reselling or embedding
Modification	Limits on derivatives	Fine-tuning and sharing

What restrictions appear most often?

Prohibited use cases, usually naming categories the provider will not be associated with. Commercial thresholds, where free use ends above a user or revenue level. Attribution requirements. Restrictions on using outputs to train competing models. And pass-through obligations that bind your own customers to terms they may not expect.

Each of these is enforceable and each is discovered late in a predictable pattern: at the moment the product does the thing the term restricts.

Who owns outputs?

Less settled than most teams assume. Provider terms generally assign whatever rights exist in outputs to the customer, which resolves the contractual question between you and the provider.

Whether outputs attract copyright at all, and who would hold it, varies by jurisdiction and continues to develop. For most commercial uses the contractual position is sufficient; for anything where exclusive ownership matters commercially, the question deserves specific advice.

What are scale thresholds?

Terms that change above a stated level of users or revenue, typically requiring a separate commercial agreement. A product built freely during development and early adoption may need a renegotiation exactly when it succeeds.

Knowing the threshold early lets it be planned for. Discovering it after crossing it is a conversation held from a weaker position.

What changes when a licence is revised?

Your position, potentially, depending on the version you accepted and whether the terms bind you to updates. Providers revise licences, and an organisation that cannot state which version it accepted for which model version cannot answer a compliance question.

Recording the licence version alongside the model version in your inventory is a small discipline that makes this answerable. See what is an ai inventory.

What should be checked before building?

Commercial use permission at your expected scale, prohibited use cases against your roadmap rather than only your current product, output restrictions if you plan any training, attribution obligations, and pass-through terms if you embed or resell.

That check takes an hour and prevents the category of problem that surfaces after the system is load-bearing.

What should you do first?

List the models in production and check whether anyone recorded the licence terms accepted. In most organisations the answer is no, and the exercise of finding out frequently surfaces a restriction that affects current plans.

Who should review them?

Legal, with engineering input on what the system actually does. A licence reviewed by engineering alone misses the obligations; one reviewed by legal alone misses whether the planned architecture triggers them.

What about indemnities?

Several providers now offer indemnification against certain intellectual property claims arising from outputs, subject to conditions such as using their safety features. Those conditions matter, and an indemnity relied upon without meeting them is not an indemnity. Reading what the protection actually requires is worth the time.

How do licences differ across deployment modes?

Considerably. The same model may carry different terms when accessed through a hosted API, downloaded for self-hosting, or obtained through a cloud marketplace, and organisations using more than one route need to track each separately.

Marketplace terms in particular are frequently assumed to match the provider's direct terms and frequently do not, because the marketplace operator adds its own conditions on top.

What happens at renewal or migration?

Terms are re-examined, and they may have changed. A model version upgrade can bring different conditions, and migrating between providers means the new terms apply to everything built on them.

Treating a licence check as part of any model change, rather than as a one-time procurement step, is what keeps the position current. It takes minutes and prevents the discovery that a restriction arrived with an upgrade nobody read.

How FISTA Solutions helps

FISTA Solutions checks licence terms against the roadmap rather than the current product, records licence and model versions in the AI inventory, flags scale thresholds before they are crossed, and raises output and pass-through restrictions where they affect product plans, through AI enablement, AI agents, and forward deployed engineers. The record behind the approach is 150+ projects for 50+ companies across 12+ countries.

To build on models without licence surprises, message FISTA on WhatsApp, or read what is open-weight ai.