[License-review] For Approval: OpenMDW License Agreement, versions 1.1 (OpenMDW-1.1)

[Richard Fontana]

On Fri, Aug 14, 2026 at 1:09 PM Michael Dolan
<mdolan at linuxfoundation.org> wrote:
>
> I welcome the feedback. I am not aware of any other AI model-focused license having been submitted previously,

There actually has been at least one, but I think OpenMDW (considering
both versions) is probably the only one seeing any real-world use.

> and I apologize if we are pushing boundaries that OSI or this committee is not prepared to cover. I had heard rumors there was interest within OSI in seeing the OpenMDW license submitted, which is part of why we brought OpenMDW here rather than treating it as outside scope. If OSI is the wrong forum, I will take that as a fair response.

Given that OpenMDW is designed to be applied to software as well as
other kinds of "Model Materials", and I suppose given the approach to
licensing taken by the OSI in the OSAID (though I assume no model
released under OpenMDW would qualify as "open source AI" under the
OSAID), I think the submission to OSI is completely appropriate, FWIW.
Given that OpenMDW-1.0 is being used already and doesn't have the
'copyright assertion defensive termination' feature of 1.1, maybe it
would be a good idea for OSI to consider 1.0 too, but that's up to the
OSI.

> The feedback seems to focus on the termination provision. I'll repaste the OpenMDW-1.1 [5] termination provision here so everyone can easily take another look at it:
>
>
> 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.
[ . . . ]
> 1. It helps to separate the termination trigger from the scope of what is terminated.
>
> On the remedy side, terminating copyright grants is not new for OSI-approved licenses.

Yes, this now seems quite clear.

> So full termination of all grants, copyright included, has a history of precedent within OSI. I suspect what is unique in OpenMDW-1.1 [5] is copyright as a trigger for termination, and I will address that in point 3. below.
[ . . . ]
> The consistent point seems to be the trigger: termination should be limited to claims about the licensed work itself. The OpenMDW-1.1 is likewise scoped to trigger on claims against the Model Materials, or in other words, the covered work.

I want to push back a little on this, though I think it may expose a
problem that may be present in existing OSI-approved licenses with
defensive termination provisions.

How do you know what the "covered work" is? This question relates to
my suggestion that there may be an OSD 9 problem.

""Model Materials" means the materials provided to
you under this agreement, consisting of: (1) one or more machine learning
models (including architecture and parameters); and (2) all related artifacts
(including associated data, documentation and software) that are provided to
you hereunder."

I've suggested there may be an OSD 9 problem because the license is
designed to apply to a bunch of different classes of materials that --
I would contend -- are of types that at least in the Linux ecosystem
(whose traditions I feel most familiar with) would have historically
been thought of as multiple, potentially differently-licensed works.
With OpenMDW, the termination provision applies seamlessly across the
whole set -- for policy reasons that are actually unclear to me other
than drafting simplicity. I suppose we can say that the Apache-2.0
termination provision has a corresponding problem because of its use
of the "Work" term, which on some level is ambiguous (""Work" shall
mean the work of authorship, whether in Source or Object form, made
available under the License, as indicated by a copyright notice that
is included in or attached to the work"). But at least as a matter of
tradition, I *think* we know that the Apache-2.0 patent termination
provision doesn't extend across, say, different Apache-licensed
projects even where released by the same licensor, perhaps on the same
website, and even where one might be a dependency for another for a
built project. For example, surely the ASF and its contributors would
not contend that asserting a patent against Apache Commons Lang leads
to termination of patent licenses granted for Apache Kafka (er,
right?).

This is where the widespread annoying practice of AI folks dividing up
materials across differently-hosted repositories, which I referred to
previously, becomes relevant. If an OpenMDW licensor releases some
Model Materials on Hugging Face and other Model Materials on GitHub,
alll associated with one specific model release, perhaps even with the
respective repositories having dissimilar names and non-obvious common
maintenance, and all that material is released under OpenMDW-1.1 -- a
situation that for all I know is already happening -- what's the
covered work in that case? Or, let's put aside the split-hosting
problem. Model Materials explicitly can be two or more models. That
suggests that those multiple models might be released in different
repositories (and perhaps other Model Materials will be released in
still other multiple repositories). The multiple models could in
principle be radically different in type. How does any licensee have
reasonable notice of what the scope of the "covered work" is? The safe
conclusion might be that anything that a licensor ever releases under
OpenMDW-1.1 constitutes one gigantic, and perhaps ever expanding,
covered work, with significant consequences for licensees thinking
about the risks associated with the termination provision.

Maybe that doesn't matter for OpenMDW-1.0, because of an uncodified
"software patents are such bad things that we can go really far in
disincentivizing patent assertions" principle, but OpenMDW-1.1 has to
be analyzed differently because it reaches licensee copyright claims.

Perhaps OpenMDW could be revised in such a way that the licensor would
have the expectation/obligation of listing out what they consider the
elements of the Model Materials to be?

> Model publishers are facing a very different economic equation than circa 1990-2005 software developers. 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. And that was the clear feedback received after OpenMDW-1.0's release.

This makes sense and I addressed the basic sociopolitical problem with
this in other responses. I would note for this response that this
doesn't explain why the termination has to be structured to apply
across the set of Model Materials. Couldn't you have it scoped to the
model parameters? Realistically, that's what we're talking about for a
copyright claim, I think.

> 5. The practical reality.
>
>
>
> The reality is that we’ve seen a marked increase in bespoke model licenses designed to cover the gaps identified in my submission. The NVIDIA Open Model License [11] terminates on copyright or patent litigation alleging the model infringes and was the inspiration behind the change in OpenMDW-1.1. The current Nemotron license [12] extends its trigger to claims that outputs infringe. The Llama family [13] (while admittedly not OSI-approved) terminates on IP claims broadly (patent and copyright, claims and counterclaims, outputs and weights), plus acceptable use restrictions. These are terms important to model builders and publishers. The trend has been that the valuable models people want to use are often not under a license that would meet OSI’s standards.

This is an important argument but I understand it as: To get model
publishers to use licenses that are closer to traditional FOSS
licenses, we need to bend the norms a little bit to accommodate their
idiosyncratic policy preferences, and if necessary modify how we
understand the OSD/'open source' in the process so that "closer to
open" then becomes "open". I guess I would just note that, while
NVIDIA, or whoever at NVIDIA makes decisions about model licensing,
seems committed to not using existing OSI approved licenses for
models, plenty of other model publishers (including recently Meta)
have been using Apache-2.0 and MIT.

Richard