Open Model Agreement

An AI model is a very large file containing learned numerical values. Whoever possesses this file can run the program on their own computer. That’s why some companies release their models for download. An Open Model Agreement is the contract text that accompanies such a release. It states what one may and may not do with the file. Users typically accept this text with a single click before the download starts.

Why “open” here doesn’t automatically mean “free”

In the software world, the term Open Source has existed for decades. It means: the source code is viewable, and anyone may use, modify, and redistribute it without restriction. Classic open-source licenses like MIT or Apache 2.0 don’t prohibit any particular use cases. Many Open Model Agreements, however, do exactly that.

Clauses against certain uses are typical: no weapons development, no mass surveillance, no targeted disinformation. Some contracts also include a revenue or user threshold. Meta's Llama license, for example, requires separate permission if a service has more than 700 million monthly users. This practically only affects a handful of large corporations, but it is legally binding.

For you as a student or for a small startup, this changes little. Nevertheless, the distinction matters because companies like to talk about “Open Source” in press releases, even though their contract imposes conditions. Experts therefore prefer to call such models “open weights” rather than open source. The weights are precisely those learned numerical values in the model file.

What such a contract typically contains

The first part specifies what is actually being shared. Almost always, this is only the finished weights. The training data and the exact training code usually remain secret. So you get the result, not the recipe.

The second part regulates permitted use. It states whether one may build commercial products and whether one may further train the model. The contract often requires attribution, such as the addition “Built with Llama” in the product name. Another point concerns derived models: whoever fine-tunes a released model must generally redistribute the result under the same conditions.

The third part concerns liability law. The provider practically excludes all responsibility for damages. If a model gives incorrect medical advice, it is the operator of the application who is liable, not the manufacturer of the model. This shift in responsibility is one reason why companies have their legal department review such contracts before deploying a model.

Llama, Gemma, and the debate over open weights

The best-known examples come from major tech corporations. Meta releases its Llama models under its own community license. Google uses similar terms of use for its Gemma models. Other providers, such as Mistral from France, rely on genuine open-source licenses for some of their models and on their own contracts for others.

In news reports, the term usually surfaces when disputes arise. The Open Source Initiative, an organization that stewards the term Open Source, has publicly denied Meta’s Llama license the designation Open Source. Such disputes are more than semantic quibbling. Authorities and universities are sometimes only permitted to use genuine open-source software.

The European AI regulation, the AI Act, also ties certain relief provisions to freely licensed models. Whether a particular Open Model Agreement fulfills these requirements is legally disputed. For you, this mainly means: if a company advertises a model as “open,” it’s worth taking a look at the license text. It will reveal whether this is a gift or a loan with conditions attached.