This term has no legal definition. A model can be announced as open and prohibit commercial use, or limit the number of users, or reserve certain uses. Between a model under an MIT licence and a model under a proprietary licence, the gap in freedom is considerable, and the difference is not visible in the announcements.
We talk about open models every week. Here is how to read what lies behind the label, because the question arises as soon as you build something serious.
The three families
Permissive licences. MIT and Apache 2.0 are the most common. They allow just about anything: commercial use, modification, redistribution, integration into a closed product. Obligations are generally limited to retaining the licence notice.
The difference between the two is modest but matters for companies: Apache 2.0 includes an explicit patent clause, which protects the user against an infringement action from contributors. MIT is shorter and silent on this point. That is why many legal departments prefer Apache.
It is under these licences that GLM-5.3-Flash in MIT and Qwen 3.8-27B in Apache 2.0 are released.
Reciprocal licences. They allow use on condition that modifications are also published under the same licence. Common in software, rarer for models.
Restrictive proprietary licences. This is where the traps lie. A lab writes its own licence, which may prohibit commercial use, cap the number of users, prohibit training another model on the outputs, or reserve certain fields of application.
Commercial use: authorised or reserved for research? This is the most frequent question and the most costly to discover late.
User threshold: some licences require negotiation beyond a given volume.
Distillation: many prohibit using outputs to train a competing model, which we touched on in our article on distillation.
Prohibited uses: a list of excluded domains, often broadly worded and uncertain in application.
What the licence does not cover
Two important blind spots, and they regularly catch people out.
Training data. A permissive licence on the weights says nothing about the legality of the data used to produce them. If litigation arises on this point, as in the publishers' lawsuits, the model's licence does not protect you.
What the model produces. The licence covers the weights, not the outputs. The question of who owns a generated text falls under a different regime, which we covered in a dedicated article.
What to check before building
Three simple and effective habits.
Read the licence, not the announcement. It is on the model's page. It is usually short, and the gap with the press release is sometimes considerable.
Archive the version you use. A licence can change for subsequent versions. The one under which you downloaded continues to apply to that specific version, and that is an argument you will only have if you kept a record of it.
Keep a copy of the weights. A downloaded file does not disappear, unlike a page that can be taken down. This has become a basic precaution since the change of owner of the main platform.
What to remember
The licence is probably the most important and least examined criterion in choosing an open model. People compare scores for hours, but do not read the paragraph that determines what they are allowed to do with it.
And it is the one that cannot be fixed later. A slightly less capable model can be replaced; a product built on a model whose licence prohibits your use has to be rebuilt.