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

[Michael Dolan]

On Wed, Aug 26, 2026 at 7:59 PM Pamela Chestek <pamela at chesteklegal.com>
wrote:

> The developer may just want to run the model, but with this license they
> have to give up any infringement claim in return. Your solution, with its
> prospective prohibition, is a much blunter instrument than examining the
> facts of the case and evaluating whether the plaintiff is acting in an
> inequitable way. Here, the FOSS developer may be acting perfectly equitably
> but still couldn't bring a claim. So it may be an over-correction for the
> problem it is trying to solve.
>
> At the end of the day I think we disagree on whose interests are more
> important to protect and it's up to OSI to decide.
>
Regarding whose interests to protect, I struggle with an assertion that OSI
is mandated to protect any particular side with more favorable terms... In
the end, we're protecting the right to use, modify, and distribute
intellectual property - all of which are enabled by the OpenMDW-1.1. I
think there may be too many presumptions clouding this thread. FOSS
developers are now or will be on both sides of the dockets McCoy linked to.
That link shows over one hundred training suits in the US alone, and the
top plaintiffs are the largest copyright holders in the world: news
organizations, studios, music labels, stock photo companies. They will sue
developers and the companies they develop for too if given the opening, all
while using models developers produce. I sense a presumption in parts of
this thread that model publishers are solely large technology companies,
and claimants are individual developers - that’s not the reality I see.

If OpenMDW-1.1 were widely adopted... I want to flag a factual premise in
the hypothetical, because it flips the conclusion. In a world of
proprietary licenses and OpenMDW, the proprietary alternative does not
preserve the developer's claim. The NVIDIA Open Model License terminates on
copyright or patent litigation. The Llama license terminates on any IP
claim, including claims about outputs, and adds use restrictions as well.
Most bespoke model licenses with substantial real-world use that I am aware
of are broader than OpenMDW-1.1 on this point - except for maybe RAIL-M,
which terminates solely on patent infringement. So I don't think a
developer would be driven to a proprietary model to preserve a claim; the
proprietary model is (I think) almost always worse for them. The
claim-preserving alternative is to use a model under Apache-2.0 or MIT,
which OpenMDW does not displace or discourage. If OpenMDW succeeds, the
ecosystem contains more open models with narrower termination terms than
the bespoke licenses they replace, not fewer.

On your blunter instrument point, Apache-2.0 does not ask whether the
patent plaintiff behaved equitably before its patent licenses terminate.
GPL-2.0 section 4 terminates automatically on any violation, however
innocent. The community chose bright-line rules because licensing at scale
cannot run on after-the-fact equity (which is inevitably subjective). A
rule everyone can read in advance is fairer than an equity standard applied
later. The bluntness is symmetrical, knowable in advance, and it is how
every defensive termination provision before this one has worked. Some
users will decide they are comfortable with using a model under those
terms, and some will decide they aren’t, just as is the case today with the
range of OSI-approved licenses.

The reverse configuration is just as real. There are small teams,
communities, and nonprofits that want to publish models openly (whether
trained from scratch, or improvements on other existing models) and cannot
underwrite litigation from well-resourced plaintiffs who could use their
model while suing over it. We have already seen a well-intentioned,
copyright-respecting model publisher decline to move forward solely because
of this exposure. Additionally, the fact that FOSS code was used in
training does not mean every developer holds a claim against every model.
McCoy said it already but, nobody knows yet. That uncertainty is precisely
what a small publisher cannot underwrite, and it is why the realistic
alternative to OpenMDW-1.1 is not Apache-2.0. It is not publishing openly
at all. I have real examples of this.

Suppose the OSI decided to publish a model trained on public domain
materials, and a publishing company disputes the provenance of some of
those works and sues, as some certainly would. Under OpenMDW-1.1, that
publisher can sue the OSI with every remedy intact. The one thing it cannot
do is keep using the OSI's model while telling a court that the model is
unlawful. I do not think the ecosystem is better served by a rule where it
can do both. Companies on both sides of use and publishing have run this
calculation as well.

On NVIDIA. I cited its license as the structural origin of the copyright
and patent trigger, not as a policy endorsement, and OpenMDW-1.1 adopted a
narrower version.

Mike
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.opensource.org/pipermail/license-review_lists.opensource.org/attachments/20260828/204620de/attachment.htm>