Providers of general purpose AI models released under a free and open source licence that allows access, usage, modification and distribution, and whose parameters including weights, model architecture information and model usage information are publicly available, are relieved from two obligations in Article 53(1): the technical documentation duty and the information to downstream providers duty. Article 54(6) gives parallel relief from the authorised representative duty on similar conditions. The relief does not apply to general purpose models with systemic risk, and the copyright policy duty and the training content summary duty in Article 53(1)(c) and (d) remain. The relief in Article 53(2) and 54(6) is narrower than the scope exclusion in Article 2(12) for AI systems, and the two provisions must not be confused.
How does the general purpose model regime differ from the AI system regime?
General purpose AI models are regulated in Chapter V, separately from AI systems. A general purpose model is defined in Article 3 by reference to significant generality and the capability to competently perform a wide range of distinct tasks, with integration into a variety of downstream systems or applications. A provider of such a model that places it on the Union market must meet the obligations in Article 53, and a provider of a model with systemic risk must meet the additional obligations in Article 55. Those obligations apply to the model provider, even where the model is later integrated into an AI system by a downstream provider. The downstream provider then has its own duties for the system. Article 2(12), which excludes certain AI systems released under open source licences from the Regulation, does not give the same exclusion to general purpose model providers. The model provider must therefore examine Articles 53 and 54, not Article 2(12), for the relief that applies to its model. A team that treats a model as an AI system for this purpose, or an AI system as a model, will apply the wrong provision and will miss duties that apply to the other category.
What conditions must an open source model meet for Article 53(2)?
Article 53(2) disapplies the obligations in Article 53(1)(a) and (b) for providers of AI models released under a free and open source licence that allows for access, usage, modification and distribution of the model, and whose parameters, including the weights, the information on the model architecture, and the information on model usage, are made publicly available. Each element of that test must be met. The licence must allow the four activities named. The parameters including weights must be public. The architecture information must be public. The usage information must be public. A model released with weights withheld, with architecture information withheld, or under a licence that restricts modification or distribution, does not meet the test on the face of the provision. A model released under a licence that allows access but not commercial use, or that imposes a user cap, must be examined with particular care, because the restriction may mean that the licence does not allow the activities named in the provision. The release evidence should record the licence text, the location of the weights, the architecture information and the usage information, with dates and versions, so that a reviewer can verify each element without needing to ask the provider for further material.
For the broader regulatory framework, see SMEs, startups and open source overview.
What duties remain for an open source model provider?
Even where Article 53(2) applies, the provider must put in place a policy to comply with Union law on copyright and related rights, including identifying and complying with reservations of rights expressed under Directive (EU) 2019/790, under Article 53(1)(c). The provider must also draw up and make publicly available a sufficiently detailed summary about the content used for training, according to the template provided by the AI Office, under Article 53(1)(d). Those duties apply to all general purpose model providers, including open source providers, except as the Act provides otherwise. The provider must also cooperate as necessary with the Commission and national competent authorities under Article 53(3), and must meet the classification and notification duties in Articles 51 and 52 where the model may present systemic risk. The provider of a model with systemic risk must meet Article 55, including model evaluation, adversarial testing, systemic risk assessment and mitigation, serious incident tracking and reporting, and cybersecurity protection, and cannot rely on Article 53(2) to avoid those duties. A provider that treats the open source relief as removing all Chapter V duties will therefore miss the copyright, summary, cooperation and systemic risk duties that remain.
What is the Article 54(6) authorised representative relief?
Article 54 requires providers of general purpose AI models established outside the Union to appoint an authorised representative established in the Union before placing the model on the market. Article 54(6) provides, in substance, that the authorised representative obligation does not apply to providers of general purpose models released under a free and open source licence on conditions similar to those in Article 53(2), except for models with systemic risk. The conditions include the licence allowing access, usage, modification and distribution and the public availability of parameters, architecture and usage information. A non Union provider that relies on Article 54(6) should record the same release evidence as for Article 53(2), with the licence and the public availability evidence, and should check the consolidated Article 54 text for the full conditions. The relief is from the authorised representative duty, not from the other Chapter V duties, and it does not remove the need for a contact point, for cooperation with authorities, or for the copyright and summary duties. A provider that has no Union presence and no authorised representative should be particularly careful to record how authorities and downstream providers can contact it, and how it will cooperate if asked.
How do codes of practice interact with open source models?
Article 56 provides for codes of practice for general purpose models, and Article 53(4) allows providers to rely on codes of practice to demonstrate compliance with Article 53(1) until a harmonised standard is published. The General Purpose AI Code of Practice, published by the AI Office in July 2025, is the voluntary instrument for that purpose, and adherence to an approved code can support a presumption of compliance to the extent the code covers the obligations. An open source provider can use the code in the same way as any other provider, for the duties that remain, but the code does not change the Article 53(2) conditions. A provider should not present code adherence as a substitute for the licence and public availability evidence, and should not present the open source relief as a substitute for code adherence where the code is relied on for a remaining duty. The record should state separately: the licence and public availability evidence for Article 53(2) and 54(6), the copyright policy, the training summary, and the code or standard relied on for each remaining duty. Keeping those strands separate makes the file easier to review and easier to update when the code, the standard or the model changes.
Continue with open source AI systems.
What does this guide not decide?
This guide does not decide whether any named model is a general purpose AI model, whether it presents systemic risk, whether its licence meets the Article 53(2) test, or whether its weights, architecture and usage information are publicly available in the sense required. Those questions need the model, the licence, the release evidence and the training compute and capability evidence, checked against Articles 51 to 55 and the Commission guidelines on general purpose models. A model provider with a specific model should confirm the classification, the systemic risk position and the open source relief with a competent authority or legal adviser, and should not rely on a repository label or on a general statement about open source as a determination.
Frequently asked questions
Does Article 2(12) exempt open source general purpose models?
No, not in the same way. Article 2(12) is a scope exclusion for AI systems. General purpose model providers must examine Articles 53(2) and 54(6), which give narrower relief from defined documentation, information and authorised representative duties, subject to licence and public availability conditions and excluding models with systemic risk.
What must be public for Article 53(2) relief?
The parameters including the weights, the information on the model architecture and the information on model usage must be made publicly available, and the licence must allow access, usage, modification and distribution. Each element must be evidenced with the licence text and the locations of the material, with dates and versions.
Does open source relief remove the copyright and training summary duties?
No. The copyright policy duty in Article 53(1)(c) and the training content summary duty in Article 53(1)(d) remain for open source model providers. The relief in Article 53(2) covers only the duties in Article 53(1)(a) and (b), and only where its conditions are met and the model does not present systemic risk.
Can an open source model with systemic risk use the relief?
No. Article 53(2) and Article 54(6) both exclude general purpose models with systemic risk. A provider of such a model must meet Article 55, including evaluation, adversarial testing, systemic risk assessment and mitigation, incident reporting and cybersecurity duties, and must meet the classification and notification duties in Articles 51 and 52.
Continue through this AI Act cluster
This article is an informational overview, not a legal determination. Check the current legal text and official implementation guidance for your system and role.
Consolidated Regulation (EU) 2024/1689 ↗AI Act Service Desk timeline ↗