A provider of a general-purpose AI model under the EU AI Act must give downstream builders technical documentation and usable information about the model's capabilities, limits and intended integration, so they can meet their own obligations. The provider must also keep that documentation current and make a copyright policy and a summary of training content available.
Key takeaways
- The model provider's duties to downstream builders are set out in Article 53 of the EU AI Act.
- Documentation handed over at integration is only useful if it describes the specific model version you run.
- Most handover failures come from silent version changes and from documentation written only for a general audience.
- Builders should collect, pin and log the upstream documents before writing integration code.
What does the AI Act require a general-purpose AI model provider to pass on?
The provider must give downstream providers enough information and documentation to understand the model and to meet their own obligations. These duties sit in Article 53 of Regulation (EU) 2024/1689, the AI Act, which applies to providers of general-purpose AI models and has applied to them since 2 August 2025.
The handover has three practical parts: technical documentation the provider keeps up to date, information for downstream providers who build systems on the model, and public materials such as a copyright policy and a training-content summary. Each part serves a different reader. The technical file serves the authorities and your compliance team, while the downstream information is what your engineers use to design the system.
Some models carry additional duties. The regulation presumes systemic risk for models trained with a cumulative compute above 10^25 floating point operations, as set out in Article 51 of the same regulation, and such providers face stricter evaluation and reporting duties. For a builder, that matters because a model under systemic-risk duties usually comes with more documentation, not less.
This article is general information and not legal or tax advice.
Who counts as a downstream provider?
A downstream provider is a company that builds an AI system on a general-purpose model it did not develop itself. The role depends on what you place on the market or put into service, not on the size of your project. The definitions and roles are set out in the text of the AI Act published in the Official Journal.
Calling a model through an API, fine-tuning it or wrapping it in a workflow can all make you the provider of a downstream system. Each path brings different documentation needs. A team that only sends prompts to a hosted model has fewer obligations than a team that modifies weights and sells the result.
What should the handover documentation contain?
The documentation should describe what the model can and cannot do, the intended uses, the input and output formats, and the known failure modes, in enough detail to test your system against them. Annex XII of the AI Act lists the minimum elements of the information the provider must give downstream providers, and a useful handover follows that list closely. The technical documentation the provider keeps for the AI Office and national authorities is a separate list, in Annex XI.
Use the following checklist when you review what a provider sends:
- The exact model name, version and release date, so you can match documents to the model you call.
- Stated capabilities, limits and known failure modes, including languages and input lengths the provider tested.
- Intended and prohibited uses, with any restrictions that affect your sector.
- Input and output formats, rate limits and the behaviour of the interface you depend on.
- A copyright policy and a public summary of the training content.
- A channel for change notices and a deprecation schedule.
Missing items are not always a refusal. Sometimes a provider has the information but has not published it in a form your team can use, so ask for it in writing.
Where does the handover usually fail?
The handover usually fails in three places: version changes that arrive without notice, documentation that stops at a marketing summary, and documents that nobody in your organisation reads before go-live. Each failure is easy to miss because the system keeps working in ordinary tests.
Version drift
A hosted model can change behaviour when the provider updates it, even when the name stays the same. If your documents describe one version and your system calls another, the documentation no longer describes what runs in production. Pinning a version where the provider allows it reduces the risk, but it also means you must plan the upgrade yourself.
Trade-offs between hosted and self-hosted models
A hosted model gives you less control over the weights and the update schedule, but the provider carries more of the operating burden. A self-hosted open-weight model gives you control over versions, yet you then hold more of the documentation and monitoring work. Neither option removes the handover; it just moves who holds the file.
Hypothetical examples of a model handover
Here are two hypothetical scenarios:
In the first, a team builds an internal contract-review tool on a hosted model. The provider's documentation states a maximum input length and warns about weaker performance on scanned documents. The team tests long contracts and scanned files before launch, and it adds a rule that routes those cases to a human reviewer.
In the second, a company runs a customer support assistant on a model whose provider releases an update without a change notice. The answer style shifts and some refusals change. The team had not pinned the version or logged outputs, so it cannot easily show what changed. It then pins the version, starts logging outputs with the model name, and asks the provider for a notice channel.
How can a builder prepare before integrating?
A builder should collect the upstream documents, pin the model version and write a short handover register before writing integration code. These three steps cost little time and make later audits and incident reviews far easier.
- Request the documentation, the training-content summary and the copyright policy in writing, and record the date you received each one.
- Pin the exact model version in your configuration and document the reason for that choice.
- Write a handover register that lists each upstream document, its version, its owner and the review date.
- Test the system against the failure modes the provider lists, and log outputs with the model name.
- Agree on a change-notice route and a deprecation plan before you go live.
Teams that need help with this work often combine the review with engineering support. The AI Act technical measures service covers implementation work of this kind, and the AI consulting team can help map the handover to your systems. Connecting logging and version records to existing systems is covered by the integrations service, and the security and GDPR guides cover the data-handling side of the same work.
Frequently Asked Questions
Does the AI Act require a model provider to share its training data?
No, the regulation does not require a provider to hand over its training datasets, but it does require a public summary of the training content that is detailed enough to show what was used. The provider must also have a policy for complying with EU copyright law. Downstream builders should read that summary before they rely on the model for regulated decisions.
Is a downstream provider responsible for the provider's documentation?
A downstream provider is responsible for its own system, not for the upstream model's documentation, but it cannot meet its own obligations without information from the model provider. Its compliance work depends on what the upstream provider passes on. Gaps in that handover become gaps in the downstream system's file.
What should a builder ask for before integrating a general-purpose model?
Ask for the current technical documentation, the summary of training content, the copyright policy, the usage restrictions and a contact for change notices. Request written confirmation of the model version you are integrating. Store each document with the date you received it so you can show what you relied on.
Does integrating a general-purpose model make my company its provider?
Not automatically. Whether you become the provider of a model depends on how you develop, modify and place it on the market, and a substantial change can shift those duties to you. Settle that question with qualified counsel before you ship.
Does Rui Codex give legal advice on AI Act obligations?
No, Rui Codex provides technical implementation support, not legal, tax or compliance advice, audits or certification. Its engineers can help you structure documentation logs, connect model interfaces to your systems and record version changes. For interpretation of your obligations, consult qualified counsel.
Conclusion
Building on a general-purpose model is a handover, and the quality of that handover decides how much work you carry later. Ask for the documentation, the training-content summary and the copyright policy before you write integration code, and record the model version you rely on. Keep the file current when the provider ships an update, because an old document can describe a model you no longer use. If you want an engineering review of how your systems connect to a model, request a project consultation at the Rui Codex contact page or email info@ruicodex.com to discuss your project.
Sources
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), text published in the Official Journal - EUR-Lex, Publications Office of the European Union