Fine-Tuning Under the EU AI Act: When It Makes You a Provider

Fine-tuning is continued training of an existing model on a narrower dataset to specialise it — taking a general-purpose AI (GPAI) model and teaching it your domain, your tone, or your task. Under Regulation (EU) 2024/1689 the question that matters is not whether fine-tuning is clever engineering; it is whether your fine-tune makes you a provider of the modified model, and therefore puts a slice of the Chapter V GPAI obligations on you. The short answer: sometimes. Where the modification is significant, you become the provider of the modification — and your Article 53 duties attach to that modification only, not to the whole upstream model.

What fine-tuning is — and the two ways it's classified

Start by separating two distinct classification questions, because fine-tuning sits across both:

• The GPAI-model layer (Chapter V). Does your fine-tune make you a provider of a GPAI model? This is the "downstream modification" question and is governed by the provider definition in Article 3(3) read with the recitals on modifying GPAI models.
• The AI-system layer (Article 6). When you wrap a fine-tuned model into a deployed system, is that system high-risk? That is a separate, use-based test under Article 6 and Annex III — see our classification walkthrough. A fine-tune does not move the system's tier; the system's intended purpose does.

This page is about the first question. Whether your finished product is high-risk is decided by what it is used for, not by the act of fine-tuning.

When fine-tuning makes you a provider

The Act's value-chain rule for high-risk systems is Article 25. A party who substantially modifies a high-risk AI system — or who puts its own name or trademark on one — steps into provider obligations. A parallel logic applies to GPAI models through the Chapter V recitals on downstream modification: when a downstream actor modifies a GPAI model, that actor can become the provider of the modified model.

The pivotal word is significant. Light prompt-tuning or a small adapter that nudges behaviour at the margins does not generally turn you into a model provider. A modification that materially changes the model's capabilities or general purpose does. The Commission's GPAI guidance frames this through a compute-share threshold: as a rule of thumb, a downstream fine-tune is treated as creating a new model — and the fine-tuner as the provider of that modified model — when the compute used for the modification is more than a third of the original model's training compute. Below that share, the fine-tune is generally treated as a modification too minor to flip provider status at the model layer. The threshold is guidance, not a bright line in the statute, but it is the working test downstream teams are expected to apply and document.

Do not confuse this with branding. The Article 25 name/trademark trigger concerns high-risk AI systems, not GPAI models. Selling a below-threshold fine-tune inside your own branded product makes you the provider of that AI system (Article 3(3)), with whatever system-level duties its use brings — but not the provider of a new GPAI model; the base model's provider keeps the Chapter V obligations.

What you owe — for the modification only

This is the relief most teams miss. When a fine-tune makes you a provider of the modified GPAI model, your Article 53 obligations attach to the modification, not to the entire upstream model. The Chapter V recitals are explicit that the downstream modifier's documentation and information duties cover the change it made and the compute and data used for the fine-tune — it is not expected to reconstruct or re-publish the original provider's full technical file.

Concretely, as provider of the modification you can owe:

• Updated technical documentation (Article 53(1)(a)) covering the fine-tuning — the data, process and compute you used — not the base model's full Annex XI file.
• Downstream information (Article 53(1)(b)) so anyone who builds on your version can comply.
• A copyright policy and a training-data summary (Article 53(1)(c)–(d)) for the content you added in fine-tuning.

If your fine-tune is itself open-weight under a qualifying licence, the Article 53(2) carve-out can lift the documentation and downstream-information duties for it — but the copyright policy and training-data summary for your added data survive, exactly as they do for any GPAI provider.

And the ceiling that fine-tuning cannot lift: if the base model is systemic-risk (the 10²⁵ FLOP presumption or a Commission designation under Article 51), or if your own cumulative training compute crosses that threshold, the Article 55 systemic-risk duties are in play. Crossing 10²⁵ FLOP through fine-tuning on top of a large base is conceivable for the biggest operations; calculate and document your cumulative compute.

GPAI obligations under Chapter V have applied since 2 August 2025; the Digital Omnibus deferral of the high-risk Annex III dates to 2 December 2027 did not move any Chapter V timeline.

Worked example: Lumen Health Analytics, ~70 employees, ~€12M turnover

Lumen, a Lisbon health-data company, licenses an open-weight foundation model and fine-tunes it on clinical-coding documents to build a coding-assistant feature it sells under the "Lumen Coder" name.

• Compute share. Lumen's fine-tuning run uses roughly 5% of the base model's original training compute — well under the one-third rule of thumb. On compute share alone, the modification would be too minor to flip model-provider status.
• The brand does not change that. Lumen ships "Lumen Coder" under its own trademark, which makes it the provider of the Lumen Coder AI system (Article 3(3)). It does not make Lumen a GPAI model provider: Article 25's name/trademark trigger concerns high-risk AI systems, and below the compute threshold the base model's provider keeps the Chapter V duties.
• Scope of duties. Lumen owes no Article 53 duties of its own; it records the compute share that keeps it below the threshold. Were a later, larger fine-tune to cross one-third, its Article 53 duties would attach to that modification only: documentation of the fine-tuning and the clinical data added, a copyright policy and training-data summary for that data, and downstream information for any integrator — never a re-publication of the base model provider's full technical file.
• System layer, separately. Whether "Lumen Coder" is a high-risk system turns on use under Article 6, not on the fine-tune. Clinical coding as a documentation-support tool is likely outside Annex III; a tool driving clinical decisions would need the Annex III analysis.

Verdict: Lumen is the provider of an AI system built on a GPAI model, not a GPAI model provider — its fine-tune is too small, and its trademark does not change that. Its obligations sit at the system layer, not in the inflated "you now own the whole model's compliance" fear teams often start with.

How Confir helps

Confir's rule-based engine runs the two questions in order: does your fine-tune make you a provider of the modified model (the compute-share test), and separately, is the deployed system high-risk under Article 6 — where the Article 25 name/trademark and substantial-modification triggers then decide who carries the provider duties. It returns the scoped Article 53 duties for the modification and flags any Article 55 exposure if compute crosses 10²⁵ FLOP. Same intake, same finding, with the rule that fired shown in plain English.

Frequently Asked Questions

Does fine-tuning a model automatically make me its provider?

No. Light fine-tuning, small adapters, or prompt-tuning that only nudges behaviour generally do not flip provider status at the model layer. You become provider of the modified model when the change is significant — the Commission's guidance treats a fine-tune using more than about a third of the base model's training compute as creating a new model. Putting your own brand on a below-threshold fine-tune makes you the provider of your AI system, not of a GPAI model; Article 25's name/trademark trigger concerns high-risk AI systems.

What is the compute-share threshold for fine-tuning?

As a working rule, a downstream fine-tune is treated as creating a new GPAI model — and you as its provider — when the compute used for the modification exceeds roughly one-third of the original model's training compute. It is Commission guidance rather than a hard statutory line, but it is the test downstream teams are expected to apply and to document.

If fine-tuning makes me a provider, do I owe documentation for the whole model?

No. Your Article 53 duties attach to the modification only — the data, process and compute of your fine-tune — not to the base model's entire technical file. You provide updated technical documentation and downstream information for your version, plus a copyright policy and training-data summary for the content you added. The original provider's full Annex XI file remains theirs.

Does fine-tuning change whether my system is high-risk?

No. High-risk classification is decided by the system's intended purpose under Article 6 and Annex III, not by whether a model was fine-tuned. A fine-tuned model used for recruitment screening is high-risk because of the use; the same fine-tuned model in a minimal-risk internal tool is not. Run the use-based classification separately from the provider question.

What if I fine-tune a systemic-risk model?

The base model's systemic-risk status does not transfer to you automatically, but watch your cumulative compute. If your fine-tuning pushes total training compute across the 10²⁵ FLOP threshold, or the Commission designates your modified model under Article 51, the Article 55 obligations — evaluation, adversarial testing, incident reporting to the AI Office, cybersecurity — apply to you. Calculate and record your cumulative compute before relying on being below the line.

Does the open-source carve-out apply to my fine-tune?

It can. If you release your fine-tuned model under a qualifying free and open-source licence that makes the weights, architecture and parameters publicly available, Article 53(2) lifts the technical-documentation and downstream-information duties for it. The copyright policy and training-data summary for the data you added still apply, and the carve-out never reaches a systemic-risk model.

Related guides

• Substantial modification
• Provider
• Deployer
• Model weights
• Article 25 — responsibilities along the value chain
• Article 53 — baseline GPAI obligations
• Provider vs deployer
• How to classify your AI system