The weights were public in February. The licence was permissive. Within a week the model had been benchmarked, quantised, fine-tuned and praised in three languages. And then, for the better part of a spring, almost nobody in Europe would put it behind an API.
The reason was not the model. It was the paperwork around the model, which is where most of the interesting questions about open-source AI now live.
A licence is not a permit
An open licence tells you what the authors will let you do. It says nothing about what a regulator will let you do, and the two have quietly come apart. Hosting a general-purpose model at scale now carries obligations on documentation, incident reporting and downstream use that attach to the operator, not the author.
The three cloud providers that looked hardest at the model all reached the same conclusion: the technical work was a weekend, the compliance work was a quarter, and the liability was open-ended.
Everyone wanted to run it. Nobody wanted to be the name on the form.
Who moved first
The eventual host was not a cloud company. It was a university consortium that already held a research exemption, already employed a compliance officer, and already had the habit of writing down what its systems did. The model went live on a Tuesday with a forty-page technical file that most commercial teams would have found humiliating to publish.
Within a month, two of the hesitant providers had licensed access through the consortium rather than hosting the model themselves. The paperwork, it turned out, was the product.
What this changes
The centre of gravity in open-source AI is shifting from whoever trains the model to whoever is willing to be accountable for it. That favours institutions with a legal department over collectives with a Discord server, which is roughly the opposite of what open source used to reward.
It also creates a new kind of intermediary: the compliant host, selling not compute but a completed form. Several have registered in the last quarter. None of them trains anything.
What to watch
Watch whether the technical file becomes a template other hosts copy, which would lower the barrier fast. Watch whether the authors of the next popular model publish one alongside the weights, and whether regulators accept it. And watch the consortium’s renewal date, because an exemption is not a business model.