Considering the OpenMDW license

The motivation behind this new license, Dolan explained, was that model distributions bundle together a number of different artifacts, including software, model weights, documentation, and more. Existing licenses are focused on software and were not intended for other types of artifacts; they also tend not to address the question of the output of models at all. The OpenMDW is an attempt to make it easier for distributors of open models to place everything under a single license. This 2025 blog post describes the motivations behind an earlier version of the license in more detail.

At its core, the OpenMDW is a permissive license, similar to the MIT license. There are a few differences, though. It has a clause saying explicitly that the license does not impose any sort of restrictions on any output created by the model. It casts a wider net than typical free-software licenses in that it grants use to "all copyright, patent, database, and trade secret rights" represented by the model distribution. What the granted rights actually cover, though, is somewhat fuzzy; the all-caps disclaimer section makes it clear that the distributor provides no guarantees in that regard. Indeed, the onus is put onto the user to determine whether they can, in fact, use the model at all:

YOU ARE SOLELY RESPONSIBLE FOR (1) CLEARING RIGHTS OF OTHER PERSONS THAT MAY APPLY TO THE MODEL MATERIALS OR ANY USE THEREOF, INCLUDING WITHOUT LIMITATION ANY PERSON'S COPYRIGHTS OR OTHER RIGHTS INCLUDED OR EMBODIED IN THE MODEL MATERIALS; (2) OBTAINING ANY NECESSARY CONSENTS, PERMISSIONS OR OTHER RIGHTS REQUIRED FOR ANY USE OF THE MODEL MATERIALS; OR (3) PERFORMING ANY DUE DILIGENCE OR UNDERTAKING ANY OTHER INVESTIGATIONS INTO THE MODEL MATERIALS OR ANYTHING INCORPORATED OR EMBODIED THEREIN.

Pamela Chestek questioned this text, describing that clearing as "an impossibility generally, and particularly if there is not even any disclosure of what the training materials are". She also said that this text could be read as a requirement imposed by the licensor — that if somebody is sued for a copyright violation involving the use of this model, they could, in addition to that trouble, be accused of having violated the license that allowed them to use the model in the first place.

The part of this license that has drawn the most attention, though, is this termination clause:

If you file, maintain, or voluntarily participate in a lawsuit against any person or entity asserting that the Model Materials directly or indirectly infringe any patent or copyright, then all rights and grants made to you hereunder are terminated, unless that lawsuit was in response to a corresponding lawsuit first brought against you.

Richard Fontana was concerned about the breadth of this clause, which covers both patent and copyright rights. Termination, he said, could affect far more than the specific material that is under dispute:

Not only does the license extend termination to copyright litigation, it also broadens the scope of termination by covering seemingly unrelated materials. For example, suppose I believe that an OpenMDW-1.1-licensed model infringes my copyrights. I sue the model licensor, but now my copyright and patent rights to some Python code distributed (in some sense) along with the model are terminated.

It is not uncommon for there to be multiple distributions of model weights intended for use by the same software base. If a copyright holder files a suit relating to a specific set of weights, they could lose access to all other models and the software used to run them as well. Fontana added that this termination of the license for unrelated materials could perhaps constitute a violation of section 9 of the Open Source Definition, which prohibits restrictions on unrelated software. In general, he expressed concerns that the "Model Materials" term, which is meant to represent the materials provided under the agreement, is not well defined.

Rob Landley wondered what would happen if a third party forks an OpenMDW-licensed project and incorporates some proprietary code into it; if the owner of that code then sues, would they lose access to the original project? Kevin Fleming, instead, pointed out a different potential problem: since it is (generally) impossible to know which data was used to train a model, the only way to determine whether it contains copyrighted material is by exercising the model itself. If a copyright holder sues over infringement, they will lose access to the model, which they need to produce the evidence to back up their suit. Simon Phipps said, simply, that "a license revoking Freedom 0 upon a copyright claim cannot assure software freedom".

Dolan responded that the termination clause is there to give symmetry to the license; it should not be possible to argue that a distribution is infringing while simultaneously enjoying the usage rights that distribution provides. But, he said, the legal position for model publishers is different from that of most software publishers, and that makes the addition of copyright to the termination clause necessary:

Models are built from large bodies of existing works, and that is where the model publisher's legal exposure arises. For those model publishers, the realistic assertion they face is that the licensed materials themselves infringe, and that infringement claim likely runs under copyright, not patent law. A patent-only provision in this context would replicate Apache-2.0's form while abandoning its function - there would not be symmetry.

Eric Schultz answered that the termination phrase constitutes "an amnesty for, depending on how courts rule, large scale copyright infringement by model creators". Fontana suggested restricting the termination to just the model weights at issue in this case. Chestek said that the license is not symmetric, because the producer of the model might know whether they have copied copyrighted materials, but the recipient has to give up any copyright claims without knowing, in advance, if their copyrights have been infringed.

The conversation has wound down, but without a clear outcome. It would seem that the OSI is not prepared to bless this license in its current form, but which changes would render it more acceptable have not really been specified, much less accepted by the proponents of the license.

In the end, the OpenMDW license seems to be aimed at reducing the risks associated with distributing an open (for some value of "open") model. Distributors would adopt this license in the hope of minimizing their chances of being sued for the violation of copyright or patent rights inherent in the model's training or seen explicitly in its output. A normal free-software license does not admonish the users of the covered software to ensure that said software is, itself, free of copyright problems. Use of the OpenMDW license may minimize the risk for distributors, but it also makes it clear that the use of LLMs may yet prove to be risky indeed.