What the licence tag does and does not settle

Four families, and what separates them

Permissive open-source: Apache 2.0, MIT, BSD. Commercial use, modification and redistribution allowed. Apache 2.0 adds an express patent grant and requires you to keep the notice; MIT is shorter and does neither. For a model you want to build a business on, this is the clean case. Qwen, Mistral’s open releases, Whisper, most BERT-family models and a large share of the Hub sit here.

Copyleft: GPL, AGPL, and in the model world occasionally applied to code rather than weights. These require derivative works to carry the same licence. AGPL extends that to software offered over a network, which matters for a hosted service. Uncommon for weights, common for the tooling around them, and worth reading when a serving stack you depend on carries it.

Open weights with conditions: the Llama community licence, Gemma’s terms, and several others. The weights are downloadable and the terms are not open-source by the usual definition. Typical conditions: acceptable-use policies that forbid certain applications, naming requirements for derivatives, and in Llama’s case a threshold on monthly active users above which you must ask permission. Most people never reach the threshold; you should still know it exists before you build on it.

Non-commercial and research-only: cc-by-nc-4.0, bespoke research licences. Fine for learning and for a demo, not for anything that earns money, including internally at a company. This is the trap people fall into, because the model is downloadable and works perfectly.

The inheritance rule

The one thing to carry from this lesson:

A fine-tune inherits its base model’s terms, whatever its own card says.

An adapter trained on a non-commercially licensed model is a derivative work of that model. Somebody tagging their repository apache-2.0 does not change what the base permits, and they frequently do tag it that way: sometimes carelessly, sometimes because they believe a LoRA is a separate artefact.

So the check is always two-step: read the model’s licence, then follow base_model and read that one, and keep following until you reach something with no parent. The Hub’s model tree makes this a few clicks.

This applies to data too. A model fine-tuned on outputs generated by a commercial API may be restricted by that API’s terms of service, which several providers have written to forbid using their output to train competing models. Whether such a clause is enforceable, and against whom, is not settled, which is exactly why a card should disclose it rather than stay silent.

What the licence does not cover

The training data. A permissively licensed model may have been trained on material whose status is contested. The model’s licence is a grant from whoever trained it about the weights they hold. It is not a warranty about what went into them, and almost no card claims otherwise.

The outputs. Whether text or images a model generates can be copyrighted, by whom, and whether they can infringe something in the training data, is unresolved and differs by jurisdiction. Some offices have said works without human authorship are not protected; others differ; litigation is ongoing in several countries. Anyone who tells you this is settled is describing one jurisdiction’s current position as though it were universal.

Your use case. An acceptable-use policy can forbid an application the licence would otherwise permit.

Where the disagreement actually sits

It is worth naming the shape of it rather than pretending there is an answer.

One side holds that training on lawfully accessed material is a transformative, non-expressive use, closer to reading than to copying, and that requiring permission for every work would make the technology impossible. The other holds that copying at scale for commercial gain is not excused by the output being different, and that creators whose work made the model possible are entitled to consent and compensation. Several countries have text-and-data-mining exceptions, usually with conditions such as lawful access and a right for rightsholders to opt out. Others have no such provision. Courts in more than one jurisdiction are actively deciding cases as this is written.

This course cannot resolve that and will not pretend to. Nothing here is legal advice. What it can tell you is what a practitioner should do: record what you used and under what terms, read the base model’s licence rather than the fine-tune’s tag, and ask a lawyer in your own country before you stake a business on an interpretation.

The practical checklist

Before building on a model commercially:

1. Read the licence file in the repository, not the tag on the page.
2. Follow base_model to the root and read that licence too.
3. Look for an acceptable-use policy: often a separate file or a linked page.
4. Check for user or revenue thresholds.
5. Check any datasets you fine-tune on, with the same two-step rule.
6. Write down what you found, with dates and links. Licences change between versions, and the version you downloaded is the one you agreed to.
7. For anything material, ask a lawyer. The cost of an hour of advice is small against the cost of rebuilding.

And the free path

If licence uncertainty is blocking you, there is almost always a permissively licensed alternative within a few points of quality. Apache 2.0 models exist at every size, for almost every task. Choosing one is often cheaper than resolving the question.

Do this now

Take the model you are most likely to build on, open its actual LICENSE file, and follow base_model until you reach a repository with no parent. Note every condition you find. This takes five minutes and it is the difference between a decision and an assumption.