Downstream of a Foundation Model: What the GPAI Documentation Package Gives You (and What It Doesn’t)

Since 2 August 2025 the provider of the foundation model you build on has owed you a documentation package. Read it as an integrator and it does something more useful than reassure you: it tells you exactly which compliance work it does not cover.

The EU AI Act’s obligations for general-purpose AI models came into application on 2 August 2025. Most of the commentary reads them from the provider’s side: what OpenAI, Anthropic, Google, Mistral and the open-weight houses now have to produce. If you integrate one of those models into your own product, the more interesting reading is from your side of the fence. Article 53(1)(b) requires the model provider to give you, the downstream provider, information and documentation so you can understand the model’s capabilities and limitations and meet your own obligations. That package is real, it is enforceable, and it is not a shield. Treat it as one and you will discover the gap at conformity-assessment time, which is the worst possible moment.

What the package is supposed to contain

The minimum contents are set out in Annex XII. It is deliberately narrower than Annex XI, which is the fuller technical documentation the provider keeps for the AI Office and national authorities on request. What flows down to you is the integrator’s subset:

• The tasks the model is intended to perform, and the type and nature of AI systems it can be integrated into.
• The acceptable use policies that apply to it.
• Date of release and methods of distribution.
• How the model interacts with, or can be used with, external hardware and software, and the relevant software versions.
• Architecture and number of parameters.
• The modality and format of inputs and outputs, and their maximum size.
• The licence.
• The technical means required to integrate the model, and the type and provenance of the training, testing and validation data together with the curation methodology.

Separately, under Article 53(1)(d), the provider must publish a summary of the content used to train the model, following the AI Office’s template. That summary is public and general by design. It is not part of your bilateral package and it is not a data inventory you can rely on for your own diligence.

What it genuinely gives you

Read charitably, the package is useful. It tells you the modality and size limits you must design around, the acceptable-use boundary you must not cross and must pass on to your own users, the licence terms that govern your commercial use, and the intended-task envelope the provider is prepared to stand behind. If you are choosing between models, Annex XII contents are the closest thing to a comparable datasheet the market has ever been obliged to produce. Demand them before you commit, not after. A provider who cannot supply them cleanly is telling you something about how they will behave when you have a production incident and need answers under a reporting clock.

The gaps it leaves you holding

The package describes the model in general. It says nothing about the model in your context, and that is precisely the boundary between the provider’s obligations and yours. Four gaps recur.

It is descriptive, not a warranty of fitness. Annex XII tells you what the model is and what it is intended to do. It does not certify that the model is fit for your purpose, your data, your regulated decision. The capabilities-and-limitations wording is information to help you assess fitness; the assessment stays yours.

It contains no evaluation against your intended use. If you deploy the resulting system in a high-risk context, the conformity assessment, risk management and the Annex IV technical documentation for the whole system are your deliverables. The GPAI package is an input to that file, not a substitute for it. Article 53 helps you meet your obligations; it does not discharge them.

The training-data information is thin where you most need depth. Type, provenance and curation methodology is not lineage. It will not, on its own, satisfy your GDPR lawful-basis analysis, your copyright exposure or your bias testing. Those remain your evidence to assemble.

It covers none of what you build on top. Your system prompt, retrieval layer, guardrails, output filtering and any fine-tuning are yours entirely. And fine-tuning is not a neutral act: past a certain point it can make you the provider of a modified model, with the upstream obligations attaching to you. We have written separately on when fine-tuning makes you a provider; the package you received says nothing about the model you have since created.

The open-source exemption is your problem, not theirs

If you build on an open-weight model released under a genuine free and open-source licence with public parameters and architecture, the provider is exempt from the Annex XI and Annex XII obligations entirely, unless the model carries systemic risk. Which means, for a large share of open-weight deployments, there is no package. The datasheet you were relying on does not exist, and the entirety of the documentation burden lands on you as the integrator. That is not a loophole in your favour. It is a transfer of work to your side of the ledger, and it should change how you weigh an open-weight model against a commercial one whose provider is obliged to hand you Annex XII.

What to do with the package when it arrives

Treat it as a supplier attestation, because that is what it is. File it against the model version it describes, because it is version-specific and the model under you will change. Map each Annex XII element to the corresponding gap you still have to close, and make that map the spine of your own technical file. Confirm your role before you do any of it: the whole obligation set turns on whether you are a deployer or a provider, and integrators routinely misjudge it. And write the package’s guarantees into the contract, because Annex XII is a floor set by regulation, not the ceiling of what you can demand: the attestations you flow down to a model supplier should go well past the statutory minimum.

The GPAI enforcement regime carries real weight from 2 August 2026, and models already on the market before August 2025 have until August 2027 to comply, so some packages you receive over the next two years will be incomplete or provisional. That does not move the line. The provider’s documentation tells you what the model is. Whether the system you wrapped around it is lawful is a question only your file can answer.