Foundation Model: The Common Name for What the Act Calls a GPAI Model

"Foundation model" is the term the industry and the research community use for a large, general-capability AI model — trained on broad data at scale, adaptable to a wide range of downstream tasks. The EU AI Act does not use the phrase. What the Act regulates, in Article 3, point 63 and Chapter V (Articles 51–55), is the general-purpose AI (GPAI) model. For practical purposes, the model people mean when they say "foundation model" is what the Act governs as a GPAI model: large language models, large multimodal and image-generation models, and similar.

The distinction worth holding onto from the first line: a foundation/GPAI model is a component, not a finished system, and being a GPAI model is not a risk tier. It does not, by itself, make anything "high-risk." The risk tier attaches to the system and use built on top of the model.

What the Act actually regulates: the GPAI model

Article 3, point 63 defines a general-purpose AI model as an AI model — including one 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, regardless of how it is placed on the market, and that can be integrated into a variety of downstream systems or applications. (Models used purely for research, development or prototyping before being placed on the market are carved out.)

So "foundation model" and "GPAI model" point at the same object from two vocabularies. When you read EU AI Act material, translate "foundation model" to "GPAI model" and look to Chapter V.

A related Act term is the general-purpose AI system — a system based on a GPAI model that can serve a variety of purposes. The model is the engine; the system is the vehicle. Obligations differ accordingly.

The obligations: Articles 51–55

Chapter V sets two layers of duty on GPAI model providers.

All GPAI model providers — Article 53. Draw up and keep technical documentation up to date; make information and documentation available to downstream providers who integrate the model so they can meet their own obligations; put in place a policy to comply with EU copyright law; and publish a sufficiently detailed summary of the content used to train the model. (Providers established outside the EU appoint an authorised representative under Article 54.)

Systemic-risk GPAI providers — Articles 51, 52 and 55. A GPAI model is classified as carrying systemic risk if it has high-impact capabilities. Article 51(2) sets a presumption: a model is presumed to have high-impact capabilities — and therefore systemic risk — where the cumulative amount of compute used for its training, measured in floating-point operations, is greater than 10²⁵ FLOP. Article 52 covers the notification procedure; Article 55 layers on the heavier duties — model evaluation and adversarial testing, systemic-risk assessment and mitigation, serious-incident tracking and reporting, and adequate cybersecurity.

Two precision points practitioners get wrong:

• The 10²⁵ FLOP line creates a presumption, not an automatic, immovable label — the Commission can also designate a model as systemic-risk on other grounds, and the threshold can be updated.
• Fines on GPAI model providers are imposed under Article 101 by the Commission, not under the Article 99 tiers that apply to AI-system operators. Don't apply the €35M/7% system tiers to a model-provider breach.

GPAI obligations (Chapter V) have applied since 2 August 2025 and were not deferred by the Digital Omnibus. Models already on the market before that date have until 2 August 2027 to comply (a transitional provision distinct from any high-risk deadline).

Why a foundation model is not a downstream risk tier

This is the misconception that causes the most confusion. People hear "GPT-class model" and assume it is "high-risk." It is not — "high-risk" is a classification of an AI system by its use, under Article 6 and Annex III. A GPAI model is a separate, cross-cutting category (Chapter V), not a fifth tier in the risk pyramid.

What that means in practice:

• If you build a customer-service chatbot on a foundation model, your system is most likely limited-risk (an Article 50 transparency duty to tell users they are talking to AI), not high-risk — because chatting to customers is not an Annex III use.
• If you build a CV-screening tool on the same foundation model, your system is high-risk — because recruitment screening is Annex III point 4. Same model, different system, different tier.

The model's nature never settles your system's risk tier. The use does. You classify the system you ship under Articles 5 and 6, and you separately rely on the model provider's Article 53 information to discharge your own duties — including AI literacy (Article 4) and, where relevant, transparency (Article 50).

Worked example

Sprektra Labs OÜ, a ~70-employee, ~€12M-turnover Estonian SaaS company, builds a writing-assistant feature on top of a large general-purpose model it accesses via a commercial API. The team calls it "our foundation model integration." Two questions get conflated; separate them.

Is Sprektra a GPAI model provider? No. It did not train or place the model on the market; it integrates a third party's GPAI model. Its supplier is the GPAI model provider, carrying the Article 53 duties (and Article 55 if the model is systemic-risk). Sprektra is a downstream provider entitled to receive the Article 53 information.

Is Sprektra's writing assistant "high-risk" because it sits on a powerful foundation model? No — the model's capability does not set the tier. Sprektra classifies its own system by use. A general writing aid is not an Annex III use, so the system is most likely limited-risk: Sprektra owes the Article 50 duties — making clear users are interacting with AI and, unless the feature only performs an assistive editing function, marking generated text as AI-generated in a machine-readable form (both live since 2 August 2026) — and the Article 4 AI-literacy duty for its staff. No high-risk conformity assessment arises.

Now Sprektra adds a module that scores job applicants' written answers for an HR client. That module performs an Annex III point 4 function — it is high-risk, independent of the underlying foundation model. The same engine now powers a high-risk system, and the full high-risk stack applies to that module. The model didn't change tier; the use did.

How Confir helps

Confir registers each AI system an organisation builds and classifies it under Articles 5 and 6 by use — so a system built on a foundation model is scored by what the system does, never by the model's reputation. Where an organisation integrates a GPAI model, Confir's structured assessment captures the downstream-provider position and the Article 53 information it should obtain from the model provider. The engine is deterministic and rule-based: identical intake yields an identical classification, with the rule that fired visible — so "we used a big model" never silently inflates a system's risk tier.

Frequently Asked Questions

Is "foundation model" a term in the EU AI Act?

No. The Act regulates the "general-purpose AI (GPAI) model," defined in Article 3, point 63. "Foundation model" is the common industry term for essentially the same thing — a large, broadly capable model adaptable to many tasks. When you see "foundation model" in EU AI Act discussion, read it as the GPAI model governed by Chapter V (Articles 51–55) and translate accordingly.

Does building on a foundation model make my product high-risk?

No. "High-risk" classifies an AI system by its use under Article 6 and Annex III — not by the model underneath it. A GPAI/foundation model is a cross-cutting category (Chapter V), not a risk tier. The same model can power a limited-risk chatbot and a high-risk recruitment tool; the system's use decides. Classify the system you ship under Articles 5 and 6, regardless of how capable the underlying model is.

What is the 10²⁵ FLOP threshold?

Article 51(2) sets a presumption that a GPAI model has high-impact capabilities — and therefore systemic risk — where the cumulative compute used to train it exceeds 10²⁵ floating-point operations. Systemic-risk models carry the heavier Article 55 duties (evaluation, adversarial testing, systemic-risk mitigation, incident reporting, cybersecurity). The threshold is a presumption, not the only route: the Commission can also designate a model as systemic-risk and can update the figure.

What does a GPAI model provider owe to companies that build on its model?

Under Article 53, every GPAI model provider must make technical documentation and usage information available to downstream providers integrating the model, so those providers can understand its capabilities and limitations and meet their own obligations. Providers must also keep technical documentation current, maintain a copyright-compliance policy, and publish a sufficiently detailed summary of training content. Systemic-risk model providers carry additional Article 55 duties.

Are foundation model obligations delayed like the high-risk rules?

No. Chapter V GPAI obligations have applied since 2 August 2025 and were not affected by the Digital Omnibus deferral, which concerns the high-risk regime. GPAI models already on the market before 2 August 2025 have until 2 August 2027 to comply — a transitional provision separate from the high-risk Annex III date of 2 December 2027, set by the Digital Omnibus, Regulation (EU) 2026/1744.

How are GPAI model providers fined?

Under Article 101, by the European Commission — not under the Article 99 tiers that apply to AI-system operators. This is a common slip: a GPAI model-provider breach is not penalised at the €35M/7% Article 5 tier. Keep the fining article matched to the actor: AI systems → national authorities under Article 99; GPAI model providers → Commission under Article 101; EU institutions → EDPS under Article 100.

Related guides

• General-purpose AI model — glossary
• General-purpose AI system — glossary
• GPAI model — glossary
• Systemic risk — glossary
• Floating-point operations (FLOP) — glossary
• Article 51: classification of GPAI as systemic-risk
• Article 53: obligations for GPAI providers
• Foundation model compliance under the EU AI Act