OpenMDW License

Anyone who publicly offers a program or an AI system for download must state what others are allowed to do with it. This permission is spelled out in a short standard contract called a license. The OpenMDW License is one such contract, and one that grants very extensive rights. Whoever publishes an AI system under this license allows everyone else to use it, modify it, and redistribute it – including commercially, meaning to make money with it. The abbreviation stands for “Model, Data and Weights”: that is, all the components that belong to a trained AI system. The text was issued by the Linux Foundation, a nonprofit organization that oversees collaboratively developed software.

The licensing chaos surrounding AI models

For classic software, the legal situation has been settled for decades. There are a handful of well-known licenses such as MIT or Apache, and every developer knows what they mean. For AI models, this was long different. Almost every major lab wrote its own license text, and many of them contained special rules.

One example: Meta's Llama models can be freely downloaded, but the license forbids certain uses and limits usage by very large competing companies. Such clauses mean a model isn’t truly open in the technical sense. For companies, every new special license means work for lawyers before a single line of code is even written. Especially at smaller companies, projects often fail because of legal review rather than technical issues.

The OpenMDW License is meant to end this chaos. It is deliberately designed as a shared standard that one learns once and then recognizes again and again. Another point: it covers not only the program code but also the training data and the weights, i.e. the learned numerical values inside the model. Whether such weights are even protected by copyright is legally unclear. The license sidesteps this problem by preemptively releasing all conceivable rights.

What the license text regulates

At its core is a very broad permission. The issuer grants all users the right to copy, modify, redistribute the material, and build something new from it. This applies to everything included: code, documentation, datasets, weights, and descriptions of the model architecture. Unlike with Llama, there is no list of prohibited applications.

There are only two practically relevant conditions. First, when redistributing, one must include the license text. Second, one loses all rights if one sues the issuer or other users over patents. This second rule is called a patent retaliation clause and works like a mutual safeguard: as long as no one sues, everyone may use the material.

Experts call such a license permissive. The counterpart is so-called copyleft licenses. These require that further developments be released again under the same license. OpenMDW does not require this. One may even sell a further-trained model without disclosing one’s own changes.

OpenMDW in model downloads and news coverage

Most often one encounters the license on platforms where AI models are available for download, especially on Hugging Face. There, every model has a license tag listed. If one finds “openmdw” there, one knows without further reading: use is permitted, including commercial use, without restriction of purpose.

In news reports, the term usually comes up when there is a dispute over which models are rightfully allowed to call themselves “open.” Critics refer to models with usage bans as open-washing, i.e. a misleading label of openness. OpenMDW serves as a benchmark for genuine openness in this context.

One common misconception should be clarified: the license says nothing about whether the training data is actually published as well. It only regulates what is allowed with what one receives. Many models under OpenMDW still only provide the weights, not the data. Anyone who wants to know how open a model really is must therefore also check which components are actually included in the download.