Introducing the Open-Source Model Licensed for Nobody in Particular

SATIRE — The Institute for Collaborative Openness, an organization founded in a glass meeting room with no chairs, has released its newest “open-source” language model under the Publicly Visible Community License. The weights are available for download by anyone who agrees not to download them, inspect them, run them, modify them, compare them to another model, train anything near them, or belong to an organization that has ever used electricity with commercial intent.

The release is being celebrated as a major step forward for openness, a term the Institute defines as “the condition in which the public may observe a thing from a respectful distance while lawyers keep it warm.” The repository includes a model card, a technical report, a 94-page acceptable-use policy, and a shell script called request_permission.sh, which prints “Permission request received” before securely deleting the request and all associated metadata.

Who may use the model?

According to the license, the model is available to natural persons acting solely in their capacity as non-natural persons, provided they do not use it for personal, educational, research, commercial, non-commercial, governmental, non-governmental, charitable, recreational, experimental, or computational purposes. This was intended to make access broad without allowing the model to be used in ways that might narrow access.

• Individuals may use the model if they are not individuals at the time of inference.
• Companies may use it if annual revenue is below zero dollars and no employee has ever had a lunch break.
• Academic researchers may use it after demonstrating that their university has never filed a grant application, owned a server, or experienced curiosity.
• Nonprofits may use it only if their mission statement contains no verbs.
• Governments may use it for transparency initiatives, but cannot execute the model because execution could reveal something.
• Open-source projects may integrate it as long as the integration is kept proprietary, unpublished, and physically buried.

The Institute emphasized that these provisions are not restrictions. They are guardrails designed to protect developers from the dangerous freedom of deciding whether software is useful. “Our goal is to preserve the model’s availability for future generations,” explained the release notes, “by ensuring the present generation cannot form a legally coherent relationship with it.”

The weights are open, in the architectural sense

Technically, the model weights can be downloaded. The files arrive as 183 shards, each named after a different category of prohibited user. To reconstruct the checkpoint, developers must accept an additional Cryptographic Responsibility Addendum stating that assembling the shards constitutes “a potentially commercial act of comprehension.” The tokenizer is included under a separate license that permits tokenization only of texts already known by all participants in the room.

Fine-tuning is allowed under the Community Improvement Exception, which permits modifications that are never saved, evaluated, deployed, discussed, remembered, or allowed to influence any future version of the model. Derivative works are permitted, provided they are sufficiently unlike derivatives to avoid being works. The license’s definitions section spends eleven pages distinguishing “output,” “results,” “consequences,” “downstream effects,” and “the general atmosphere after an output occurs.”

A practical guide to compliance

Software engineers considering the model should follow the recommended compliance workflow:

1. Do not clone the repository. Reading the git clone command may create an implied relationship with the weights.
2. Do not ask legal to review the license. Legal review is an organizational use, and organizations are expressly included among the things that must not use it.
3. Do not run the model locally. Local execution is particularly risky because it creates the appearance that you control a computer.
4. If you accidentally generate text, do not read it. Reading may establish that the output has utility, which is forbidden outside designated zones of non-utility.
5. If the model improves a bug report, revert the fix immediately and document that no economic value was retained.

For teams that need actual production access, the Institute offers an Enterprise Non-Use Agreement. Under this plan, customers receive a handsome PDF confirming that they have not used the model at scale. Pricing is calculated per million hypothetical tokens, with volume discounts for enterprises willing to imagine larger deployments without ever performing them.

The benchmark results remain encouragingly inaccessible

The model reportedly achieves 117.4% on the Institute’s new Responsible Non-Deployment Benchmark, which measures whether a system can decline to participate in its own evaluation. It also leads on the Legal Interpretability Suite, where models are scored on their ability to summarize a license while expressing no belief that a human being could comply with it.

Independent reproduction has not yet occurred, largely because reproduction falls under the license’s broad definition of “creating a situation in which the model exists twice.” Still, the technical report assures readers that the model is competitive with leading systems in tasks that the Institute cannot identify without accidentally enabling comparison.

Openness, clarified

There is a useful distinction underneath the joke. Publishing weights does not by itself make a model open source: practical rights to use, modify, and redistribute matter, as do license terms that allow real people and organizations to exercise those rights. Engineers evaluating a model release should read the license before planning around the word “open.”