On licensing · The But Is It Open? blog

The Clause Holding Up OpenMDW

· 5 minute read

The Linux Foundation asked the OSI to approve OpenMDW-1.1 on August 13. The usual 60-day decision date has passed with no ruling, and the review keeps returning to one sentence: a termination clause that now covers copyright as well as patents.

Abstract painting on ivory paper: layered translucent ochre planes on the left and burgundy planes on the right, separated by an open gap of bare paper, with a single small sharp navy sliver lodged at the edge of the gap holding the two sides apart.

The Clause Holding Up OpenMDW

I checked this morning to see whether the Open Source Initiative had approved OpenMDW-1.1, and I half expected to be writing about the answer. The Linux Foundation submitted the license on August 13. The OSI's review process says a decision normally comes about 60 days after submission, which means the date passed a month ago.

There's no answer yet. No recommendation from the License Committee has appeared on the review list, no board vote has been reported, and OpenMDW isn't among the approved licenses. The thread was busy for a month and then went quiet. The last message about the license was posted on September 14.

This matters more than most license reviews. OpenMDW is a permissive license written for AI models: one document that covers the weights, code, data and documentation in a release, and grants copyright, patent, database and trade secret rights together. It solves a real problem, and companies are already using it. Arcee moved its whole Trinity family onto it in May. On Hugging Face today, NVIDIA's Cosmos 3 Super and Nano carry OpenMDW-1.1, and so do the Nemotron 3.5 Lightning and Bolt checkpoints. The older Nemotron 3 Super is still under NVIDIA's own model license.

In the index behind this site, Nemotron and Cosmos 3 both score a 2 on Permission. The terms are generous. They just aren't terms the OSI has recognized, so anyone who wants to rely on them has to read a new legal text and decide for themselves. Approval would change that. It would also be the first time the OSI put its name on a license written for models instead of software.

The sticking point is one sentence

The review didn't stall on the general shape of the license. Almost the entire argument on the list is about the termination clause. If you sue anyone claiming that the released materials infringe a patent or a copyright, you lose all of your rights under the license, unless you were sued first.

Open source developers know this move from patents. Apache 2.0 has a version of it: sue claiming the software infringes your patent and your patent license from that project ends. The point is to keep patent fights away from shared code. OpenMDW-1.1 adds copyright to the trigger. Mike Dolan, who submitted the license for the Linux Foundation, told the list that this was deliberate and that it was the only substantive change from version 1.0.

The one change that turned 1.0 into 1.1 is the change holding it up.

Everyone holds a copyright

Several reviewers objected. Richard Fontana asked whether approving the clause would look like the OSI siding against developers who want to enforce their own open source licenses in court. Simon Phipps made the point I keep coming back to. Patent suits are mostly fought between companies that own patents, but every contributor to every project holds a copyright. A clause that punishes copyright claims reaches far more people, and many of them are individuals.

Models make the problem harder. They're built from enormous amounts of other people's work, and as Pamela Chestek noted, someone accepting the license usually has no way of knowing whether their own writing or code is in the training data. Dolan later clarified that a complaint about training alone wouldn't trigger termination; the suit has to claim that the released materials themselves infringe. That narrows the clause, but the person deciding whether to sue still can't see what's inside.

The defenders have good arguments. Luis Villa pointed out that MPL 2.0 already ends copyright grants when someone brings a patent claim. Josh Berkus argued that any case against a copyright trigger works against a patent trigger too, and that rejecting one would mean revisiting a lot of approved licenses. I don't think he's wrong about the logic. Where I part ways is over who gets caught: a patent clause mostly catches companies, while a copyright clause could catch a novelist or the maintainer of a small library.

What approval would settle

Fontana suggested early in the thread that the OSI could also consider OpenMDW-1.0, which is already in use and doesn't have the copyright trigger. That looks like the most practical way through. The Linux Foundation could narrow the clause, or the OSI could approve the version without it, and the bet would pay off later than people hoped.

Even then, approval would answer only one question. A license tells me what I'm allowed to do with Cosmos 3. It doesn't tell me what went into training it, and it doesn't let me rebuild it. Fontana said as much: he assumed no model released under OpenMDW would meet the OSI's definition of open source AI, which asks for more than a license. Preston Maness went further and argued that the OSI shouldn't approve any model license that doesn't require the training data.

Read the thread

The whole review is public. It begins with Dolan's submission, which includes the full license text and the Linux Foundation's case for why software licenses don't fit model releases. The replies, collected in the August and September archives, come from license lawyers and longtime open source people, most writing in a personal capacity. It's long and technical, and it's the best record I've found of people working out whether the old rules stretch to cover models.

But is it open? In substance OpenMDW is as permissive as model licenses get, and NVIDIA and Arcee deserve credit for choosing it over custom terms. The OSI just hasn't said yes yet, and the reason it hasn't comes down to a single sentence. For now it's still the bet. I'll check the list again next month.

Derived from the openness research behind But Is It Open?

← More from the blogJoin the newsletter ↗