AI Act Commission Guidelines: general-purpose AI models

The obligations of providers of general-purpose AI models under the AI Act will be applicable from 2nd August 2025. To guide providers towards compliance, the Commission launched a consultation in May 2025 on the first draft of its guidelines and published the final version of the “Guidelines on the scope of the obligations for general-purpose AI models” on the 18th of July 2025 (the ‘Guidelines’). Moreover, the Commission has received the third and final version of a much debated General-Purpose AI Code of Practice that will help providers to demonstrate compliance. General-purpose AI models (GPAI) are AI systems trained on large volumes of data that can perform a wide range of tasks and be integrated into other products and services. Under the EU AI Act, providers of GPAI models face specific transparency and documentation obligations, with stricter requirements for models that pose systemic risk. This article explores some key aspects of the recently published Guidelines: 

1. Defining a general-purpose AI model and systemic risk

A fundamental aspect of the Guidelines is to clarify the definition of a general-purpose AI model, determining which organisations will need to comply with the obligations of Chapter V of the AI Act.  

Article 3(63) AI Act defines a general-purpose AI model as an AI model “trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks […]”. This raises a core question of what constitutes (1) a sufficient level of generality; and (2) a wide range of distinctive tasks, to become within scope of the AI Act provisions.  

Recital 98 AI Act suggests that models with at least one billion parameters and trained with a large amount of data would fulfil the generality requirement. The AI’s Office preliminary approach is to set a threshold amount of computational resources used to train a model (training compute), which encompasses a number of parameters and amount of training data, rather than measuring model size and training data size separately. The AI’s Office initial approach was to presume that a model is a general-purpose AI model if its training compute is greater than 10²² FLOP, however, the final version of the Guidelines has raised this threshold to 10²³ FLOP (para 17).  

This training compute presumption is rebuttable since, in addition to the training compute, the modality of the data used for training and that the model is able to understand and generate should also be considered (para 19). The Guidelines provide examples of what should be considered a general-purpose model based on these criteria, which are not entirely clear, for example, a natural language model with 10²⁴ FLOP training on a broad range of natural language data should be considered to perform a wide range of tasks (page 8 , example 1), however, a natural language model with 10²⁴ FLOP that transcribes speech or generates speech from text, is considered to perform a narrow set of tasks, therefore, it shall not be considered a general-purpose AI model (page 8, examples 3 and 4).  

When it comes to systemic risk and in the absence of delegated acts that further specify what is considered to be ”high impact capabilities” (para 28), a general-purpose AI model is presumed to have high-impact capabilities when the cumulative amount of computation used for its training is greater than 10²⁵ FLOP and the provider must notify the Commission without delay and in any event within two weeks after that requirement is met (Article 52(1) AI Act). Providers can submit arguments to demonstrate that, exceptionally, their model does not present systemic risks despite meeting the presumption requirement, but the obligations still apply while the provider is contesting the designation as a general-purpose model with systemic risk.   

2. Providers of General-Purpose AI models: downstream modifiers

The Guidelines add a new category of operators in the AI value chain which is neither defined nor referred to in the AI Act: “downstream modifiers” defined as the entities that obtain and modify a general-purpose AI model that was originally developed by a third party. However, the key question is when the downstream modifiers will be considered providers of general-purpose AI models and assume their obligations. The Commission deems that not every modification of a general-purpose AI model will lead to the downstream modifier being considered a provider, (para 61), only if the modification leads to a significant change in the model’s generality, capabilities, or systemic risk (para 62). The Guidelines continue stating that “an indicative criterion for when a downstream modifier is considered to be the provider of a general-purpose AI model is that the training compute used for the modification is greater than a third of the training compute of the original model” (para 63). If the original training compute is not known by the downstream modifier, the threshold should be a third of the threshold for a model being presumed to be a general-purpose AI model, i.e. currently 10²³ FLOP and 10²⁵ FLOP for models with systemic risk (para 64).   

One key takeaway is that if a downstream modifier becomes a provider of a general-purpose AI model (without systemic risk), their obligations are limited to the modification conducted, i.e., documentation, copyright policy, and training content summaries. Additionally, the modifier must comply with Article 54(1) AI Act regarding authorised representatives (para 68 and 69). Nonetheless, if a downstream modifier becomes the provider of a general-purpose AI model with systemic risk, their obligations are not limited to the modification conducted. In this case, a new systemic risk assessment and mitigation must be performed for the entire modified model, taking into account any available information about the original model, and equally important, the provider must also notify the Commission (page 71).   

3. The notion of placing on the market

Providers are in scope of the AI Act if they place a model on the EU market, irrespective of their establishment or location. Providers established in third countries must appoint an authorised representative in the Union before placing the model on EU market.  

Placing on the market is defined under Article 3(9) AI Act as “the first making available of an AI system or a general-purpose AI model on the Union market,”  and Article 3(10) defines making available on the market as “the supply of an AI system or a general-purpose AI model for distribution or use on the Union market in the course of a commercial activity, whether in return for payment or free of charge”.  

The guidelines set specific examples for models to be considered ‘placed on the market’ including for those models which are integrated by a third-party into an AI system without modifying it where, in principle, the developer shall be considered the provider of the general-purpose AI model and assume the obligations in Chapter V and the downstream actor the provider of the AI system having to comply with the applicable requirements and obligations for AI systems laid down in the AI Act (para 57). 

General-purpose AI models can be placed on the market in various ways, including through software libraries, application programming interfaces (APIs), direct downloads, or as physical copies. If a provider integrates its own general-purpose AI model into its own AI system and makes it available on the market, the model is considered to be placed on the market, and both model and AI system obligations apply (Recital 97).  

4. Open-source exemptions

The AI Act generally applies to providers of general-purpose AI models released as open-source. However, recognizing the contribution of free and open-source software and data to research, innovation, and economic growth through open sharing, use, modification, and redistribution, the AI Act provides exemptions from certain obligations, unless the model is not a model with systemic risk. These exemptions include: 

• Documentation: exemption from obligations to document technical information for the AI Office, national authorities, and downstream providers. Note that the obligations to put in place a policy to comply with Union copyright law and the summary about training still remain.  

• Authorized Representative: exemption from the obligation to appoint an authorized representative for providers established in third countries.  

The Act has been highly criticised for the vagueness of the term ‘open-source’ and  for not accounting for partially open AI models (e.g. Meta’s Llama Model) (see Hence/Mustać – AI Act Compact). The guidelines clarify the requirements to abide by the open-source exemption, amongst others:   

• Rights to access, use, modify, and redistribute shall be fully granted, if one on of them is missing, the licence cannot be considered free and open (para 78).  
• Freedom to distribute onwards, often subject only to crediting authors and retaining copyright notice. Recipients should generally be able to redistribute derivatives under different licence terms (including closed-source) (para 82) 
• Restrictions that disqualify include limitations to non-commercial or research-only use, prohibitions on distribution, user scale thresholds, or requirements for separate commercial licences for specific use cases. 
• According to Recital 103, there should also be a lack of monetisation and the Guidelines list some examples of what should be considered monetisation, e.g., paid technical support or other kind of services which are mandatory to contract for model performance and security; or collection and processing of personal data unless exclusively and strictly for model security and without commercial or financial gain (para 86). 
• Information around the model’s “parameters, including the weights, the information on the model architecture and the information on model usage, are made publicly available” (Article 53.2).  

Conclusion 

The Guidelines reveal that one of the top concerns for the Commission and the AI Office is to provide mathematically measurable definitions of one of the broadest terms found in the AI Act: the generality criterion for general-purpose AI models. The threshold provided is clear (10²³ FLOP) but the second criteria, which is the ability of the model to perform a wide range of distinctive tasks, has been considered without providing clear insights, thus creating legal uncertainty. Other matters, such as the obligations for downstream modifiers, the definition of ‘placed in the market’, and clarifications on the exemptions for open-source models, have been tackled with more success. Lastly, the Commission commits to reviewing and updating these guidelines as necessary, considering practical experience, technological, societal, and market developments, including interpretations from the Court of Justice of the European Union.  

Trilateral’s Data Protection advisors and the Responsible AI  team will assist your organisation to build procedures compliant with data protection requirements while promoting responsible AI. In addition, our Trilateral GRC Suite platform equips an organization with the essential tools to showcase a strong and flourishing culture centered around data protection and responsible AI. Please contact any member of Trilateral’s team if you would like to discuss our services.   

You can also read more insights from Trilateral Research by following us on LinkedIn.