Open-Source Does Not Buy the EU Exemption It Implies

Beijing's Pitch: Open Is the Responsible Choice

Xi Jinping closed the 18th BRICS summit in New Delhi on September 13 with a five-point plan for the bloc, one of which is a BRICS AI open-source zone. "China will be a pioneer in establishing a BRICS AI open-source community, support cooperation in developing and applying large language models, hold specialized AI seminars and training courses and build an open ecosystem for AI," Xi said. The framing was explicit: Chinese labs already ship world-leading open-weight models whose code and parameters anyone can inspect and adapt, while the leading US labs keep their frontier models closed.

That framing has a real audience. An open license reads as the more accountable, more inspectable choice to a buyer who cannot audit a closed model's training data or safety testing at all. Beijing is betting that framing travels well into markets, including the EU, that have spent three years writing rules about AI transparency.

Brussels Already Answered the Question, Differently Than It Sounds

The EU AI Act's Article 53(2) does give a free and open-source model with publicly available parameters a real exemption, dropping the technical-documentation and downstream-information duties that every other General Purpose AI provider owes under Article 53(1), enforcement of which began August 2, 2026. So the instinct that open equals less friction is not wrong on its own. It is conditional in a way the BRICS pitch never mentions: the exemption holds only below the Act's systemic-risk threshold of 10 to the 25th floating-point operations of training compute. It never covers the copyright policy or the training-content summary regardless of license.

The Exemption Was Built for Small Models, Not Flagship Ones

A license does not shrink a model's training run. The systemic-risk threshold is a compute number. The frontier-scale open-weight models China is proudest of, the ones large enough to compete with closed US labs on capability, are exactly the ones most likely to sit at or above it. The moment one does, Article 53(2)'s exemption stops applying and the provider owes the full Article 53(1) package: technical documentation and a downstream information package, on top of the copyright and training-summary duties that never went away. A smaller, genuinely open hobbyist model keeps the exemption. A flagship model good enough to headline a geopolitical pitch is the one least likely to qualify for it.

This is not a loophole Brussels forgot to close. It is the opposite: the systemic-risk carve-out exists specifically so that scale, not license type, decides how much documentation a model owes the market it operates in.

What This Changes for an EU or UK Buyer

Checking a model's license before checking its training-compute footprint gets the compliance question backwards. An operator evaluating a Chinese open-weight model for the transparency Beijing is marketing should ask for the specific FLOP figure and whether the provider has made a systemic-risk determination, not treat the open license itself as proof of anything under EU law. If that figure is at or above the threshold, the buyer is looking at a vendor based outside the EU that owes full technical documentation and downstream information under a regime it has far less practical incentive to comply with than an EU-based provider does.

Beijing's pitch and Brussels' rule are not in conflict, they are just answering different questions. One is about who gets to inspect a model's weights. The other is about who has to produce paperwork proving what the model was trained on and how it behaves. A buyer who conflates the two is one systemic-risk determination away from an unpleasant compliance surprise.

Read next: Brussels Can Now Pull An AI Model From The EU Market  |  EU Regulators Can Now Fine AI Firms 7% of Revenue