What regulators ask of an open model

The rules arrived, and they are not aimed where people assume

Most writing about AI regulation talks about frontier labs. A large part of the actual text applies to anyone who places a model on the market, and fine-tuning a model and publishing it can put you in that category. It is worth twenty minutes to know where the lines are, because the compliance burden for an ordinary team is small if you know about it in advance and unpleasant if you discover it later.

Two warnings before the substance. Details here move; check the current text rather than trusting a lesson. And this is orientation, not legal advice.

The EU AI Act, in the part that touches you

The Act regulates general-purpose AI models as a category of their own, with obligations that began phasing in during 2025. For a provider of a general-purpose model, the core duties are:

• Technical documentation of the model, its training and testing, kept and made available to authorities and to downstream providers who build on it.
• A copyright policy, including respecting rights reservations expressed in a machine-readable way, which is the clause that gives robots.txt style opt-outs legal teeth inside the EU.
• A sufficiently detailed summary of the training content, published, following a template the AI Office provides.

There is a partial exemption for models released under a free and open-source licence, and three things about it are regularly misreported.

1. It does not cover the copyright policy or the training-content summary. Those two survive the exemption. So an open release still owes a public data summary.
2. It does not apply to models with systemic risk, defined with a compute threshold around ten to the power of twenty-five floating point operations of training compute. That is a frontier-scale number; ordinary fine-tuning is nowhere near it.
3. The exemption depends on the licence being free and open source, which the community licences discussed earlier may not be. A bespoke licence with use restrictions and a user cap is not obviously within it.

When a fine-tune makes you a provider

This is the part teams miss. If you take an open base model, fine-tune it, and put the result on the market under your own name, you can acquire provider obligations for the modified model. Guidance has generally focused those obligations on the modification you made rather than on the entire model, and there are thresholds below which a small tune is not treated as creating a new general-purpose model. But the direction is clear: publishing weights is an act with duties attached, and the answer is documentation you were better off writing anyway.

Separately, if you use a model in a high-risk application (employment, credit, education, essential services) deployer obligations apply regardless of whether the model is open, and those are considerably heavier.

Elsewhere

United States. No general federal statute. Sector regulators apply existing law, and several states have passed transparency and disclosure requirements. Export controls target advanced chips rather than weights, though proposals to cover model weights have been discussed repeatedly and none has settled.

China. Generative services offered to the public require filing and security assessment, and content labelling rules apply to synthetic media. This is why some Chinese-origin open models behave conservatively on certain topics: the domestic service obligations shape the tuning even where the weights themselves travel freely.

Everywhere. Content-provenance and labelling rules for synthetic media are the fastest-moving area, and they attach to the deployer more often than to the model.

A proportionate response

For a small team shipping a product on open weights, the whole of this reduces to a short routine.

• Keep the model inventory you already started: model, revision, licence, date accepted, where it runs.
• If you publish weights, publish a model card with intended use, evaluation, limitations and a description of training data. That satisfies most of what is asked and is good practice regardless.
• If you fine-tune on data you did not create, record where the data came from and under what terms.
• Decide, once, in writing, whether any of your uses fall in a high-risk category. Usually the answer is no, and having written down why is worth more than the sentence suggests.

The compute threshold is the number to remember, because it settles most anxiety. If your training run cost less than a house, you are not the model this clause was written about.