The phrase open source has a settled meaning that most model licences do not satisfy. Weights being downloadable is not the same as being permissively licensed, and several of the most widely used open-weight models ship under custom terms written by the releasing company rather than under a recognized open source licence.
For a research project this rarely matters. For a product you are selling, it matters a great deal, and it is the kind of thing that gets discovered during due diligence rather than during development.
This is general information, not legal advice. Consult a qualified professional for your situation.
What changed in 2026
- Licence diversity increased. More labs released weights under bespoke terms rather than adopting a standard licence, which made per-model review necessary rather than optional.
- Acceptable use policies gained weight. Terms incorporated by reference — and updatable independently of the licence — became a recognized compliance concern.
- Enterprise procurement started asking. Buyers began requiring model licence provenance in vendor questionnaires, which pushed the question upstream to product teams.
- Distillation clauses proliferated. Restrictions on using a model's outputs to train other models became a common and easily overlooked term.
Licence categories
| Category |
Examples of terms |
Commercial risk |
| Permissive standard licence |
Attribution, no warranty |
Low; well understood |
| Custom community licence |
User thresholds, naming requirements, acceptable use |
Medium; read carefully |
| Research-only licence |
No commercial use at all |
High; do not ship on it |
| Non-commercial share-alike |
Derivatives must match terms |
High for commercial products |
| Weights available, no licence stated |
Ambiguous |
High; ambiguity is not permission |
The last row deserves emphasis. A model uploaded with no licence file is not permissively licensed by default — absence of stated terms means absence of granted rights, not the reverse. Teams routinely treat an unlicensed upload as free to use, which is the assumption most likely to cause a problem.
The clauses that actually bind
User thresholds. Some community licences grant broad rights below a stated user count and require a separate agreement above it. If your product succeeds, you cross the line, and the time to discover that is before launch.
Attribution and naming. Requirements to display a notice, or to include the model family name in your derivative's name, are common and easy to comply with — and easy to breach by omission.
Acceptable use policies. These are typically referenced rather than reproduced, and they can be revised by the publisher after you have built on the model. Track the referenced document, not just the licence text you downloaded.
Output restrictions. Whether you may use generated outputs to train another model varies. If your roadmap includes distilling a smaller model from a larger one, this clause determines whether that is permitted.
Output ownership. Separately from what you may do with the weights, check what the licence says about ownership of what the model produces. Most permit commercial use of outputs; the terms vary and matter for anything you are licensing onward.
Weigh all of this against the alternative of an API, where terms are contractual and clearer. The self-hosting decision is usually about cost, data residency, or latency — the framing in AI inference providers compared covers the middle ground of hosted open models.
Common mistakes
- Treating open weights as open source. Different meanings, different obligations.
- Reading the hub page summary instead of the licence file. Summaries are written for convenience, not accuracy.
- Ignoring the referenced acceptable use policy. It is part of the terms and it can change.
- Missing user thresholds. Compliance at launch does not mean compliance at scale.
- Assuming fine-tuning creates a clean-room model. A derivative typically inherits the base licence's conditions.
FAQ
Can I use an open-weight model in a commercial product?
Frequently yes, subject to the specific licence. Read it, note any thresholds and attribution requirements, and keep a record of the version you relied on.
Does fine-tuning change the licence?
Generally not. A fine-tune is usually a derivative work and inherits the base model's terms unless the licence explicitly says otherwise.
What if the licence changes after I ship?
The version you accepted typically governs your existing use, but referenced policies may be dynamic. Archive the licence text and any referenced documents at the time you adopt a model.
Are permissively licensed models good enough?
Increasingly, yes. Several strong models ship under standard permissive licences, and for teams wanting to avoid this analysis entirely, restricting to those is a legitimate procurement policy.
Where to go next
For the hosting decision, read AI inference providers compared and AI compute leasing explained. For smaller self-hosted options, small language models explained.