Two years ago, the open-weights debate was whether governments would restrict them at all. That question resolved — but into two different answers.
The United States decided open weights are an instrument of influence: if the world builds on American models, American values and American chips come along. The European Union decided open weights are a product category that needs rules, albeit lighter ones than closed frontier systems. Both positions are coherent. Together they are complicated.
Why it matters
Open weights are the substrate of a large fraction of applied AI: startups fine-tune them, enterprises self-host them for data control, and researchers depend on them for reproducible science. Regulatory divergence raises the cost of serving a global market with a single model artifact — and advantages the labs large enough to maintain compliance teams per jurisdiction.
There's a strategic layer too: Chinese open models have taken significant global download share, which is precisely what US policy is trying to counter and what EU rules don't distinguish on.
How it works
Under the EU AI Act, providers of general-purpose AI models owe documentation, copyright-policy and summary obligations; models above a compute threshold deemed to pose systemic risk owe additional evaluation and reporting. Free and open-source releases get exemptions from some obligations — but not all, and the exemptions don't cover systemic-risk models. The US approach, by contrast, pairs export-control debates with an explicit policy preference for open American models, treating openness itself as the safety and competitiveness strategy.
For a developer, the practical questions are: which obligations attach to the model I fine-tuned, did I become a 'provider' by modifying it, and does my downstream product inherit documentation duties? The answers differ by jurisdiction and, in the EU, by how the model was licensed and released.
Evidence
The AI Act's general-purpose obligations began applying in 2025, with the Commission publishing guidance and codes of practice that major labs — including open-weight providers — have signed onto with varying reservations. US policy documents and the AI Action Plan explicitly endorse open-weight models as a national strategy, and American labs' release cadence of open models has accelerated in step.
Download and deployment data from model hubs shows the stakes: Chinese open models lead global downloads in several categories, which US officials cite as the competitive rationale for the pro-open stance.
The competing read
The US position holds that openness is self-regulating at scale — more eyes, more audits, more diffusion of capability away from any single actor — and that regulation of weights is unenforceable anyway. The EU position holds that capability doesn't care about license terms: a powerful model is a powerful model, and openness changes who can use it, not what it can do. Safety researchers are genuinely split between the two, which is why the divergence is likely to persist rather than converge.
What happens next
Watch the first AI Act enforcement actions involving open-weight providers, which will define how seriously the exemptions hold, and watch whether US export policy toward open weights stays permissive as model capabilities rise. The real tell will be the next generation of frontier-adjacent open releases: which jurisdiction's rules the releasing lab optimizes for says where the power sits.