The Open-Weight Exemption Under Scrutiny
The EU AI Act, which took full force for high-risk AI systems in August 2025, currently exempts general-purpose AI models released as open weights from most provider obligations — on the rationale that the releasing organisation cannot control downstream deployment. That carve-out is now under active review, according to statements from Commission officials at an AI governance summit in Brussels this week.
The proximate cause is the rapid proliferation of fine-tuned derivatives of Llama 4 and Mistral models deployed in clinical decision support, credit scoring, and automated legal document review — all domains classified as high-risk under Annex III of the Act. In these deployments, the fine-tuning organisation often lacks the resources or expertise to conduct the full conformity assessment required of high-risk AI providers.
What Revision Might Look Like
Regulators are considering three options: (1) extending light-touch obligations to fine-tuned derivatives of open models when deployed in high-risk contexts; (2) requiring platforms that host or facilitate deployment of open models in regulated sectors to act as the responsible provider; or (3) maintaining the exemption but creating a voluntary certification scheme with liability protections for compliant deployers.
AI industry groups have lobbied strongly against option 1, arguing that it would fragment the European open-source AI ecosystem. Academic researchers warn that even light obligations would chill research use of open models in clinical settings.
Timeline
The Commission is expected to publish a consultation paper by September 2026, with potential delegated act revisions entering the legislative pipeline in early 2027. No changes take effect without a further 12-month transposition period.