Fine-Tuning a GPAI Model Under the EU AI Act: When You Become the Provider

Why Fine-Tuning Raises a Classification Question

Thousands of companies adapt existing general-purpose AI models: instruction tuning on proprietary data, domain adaptation for medicine or law, reinforcement learning for product behaviour, quantisation and distillation for deployment. Regulation (EU) 2024/1689 regulates GPAI models through their providers, so the practical question is who counts as the provider of a modified model — the original developer, the fine-tuner, or both. The answer determines who owes the Article 53 duties: technical documentation, downstream information, a copyright policy and a public training data summary.

Recital 109 sets the principle: where a GPAI model is modified or fine-tuned into a new model, the obligations of the new provider should be limited to that modification or fine-tuning — for example, complementing the existing technical documentation with information on the modification, including the new training data. The regulation thus anticipates that fine-tuners can become providers, while keeping their burden proportionate.

The Commission's One-Third Compute Yardstick

The AI Act itself does not quantify when a modification creates a new model. The Commission's guidelines on the scope of GPAI obligations, published in July 2025 alongside the Code of Practice, filled the gap with an indicative threshold: a downstream modifier should be considered the provider of a new GPAI model where the compute used for the modification is greater than approximately one third of the original model's training compute. Because downstream modifiers often cannot know the original training compute, the guidelines offer practical proxies tied to the regulation's own reference points — roughly one third of 10^23 FLOPs for the question whether the modification makes the modifier a GPAI provider at all, and one third of 10^25 FLOPs in the context of the systemic-risk presumption.

The consequence of the yardstick is reassuring for most practitioners: typical enterprise fine-tuning — thousands or millions of examples on top of a model trained on trillions of tokens — consumes a tiny fraction of original training compute and does not create a new model provider. Heavy continued pre-training, by contrast, can cross the line.

If You Cross the Line: Scoped Obligations

A fine-tuner who becomes a provider of a new GPAI model owes the Article 53 duties scoped to the modification:

• Technical documentation under Annex XI covering the modification: the fine-tuning methodology, the new training data, evaluation results for the modified model, and compute used.
• Downstream information under Annex XII for those who integrate the modified model, complementing the original provider's documentation.
• A copyright policy covering the content used for the modification, including respect for text and data mining opt-outs in any new crawling.
• A public training data summary, using the AI Office template, describing the content used for the fine-tuning rather than the original pre-training corpus.

If the modified model qualifies for the open-source exemption — free licence, public weights, no monetisation, no systemic risk — the documentation duties relax in the same way they do for original providers. And if cumulative compute including the modification crosses the systemic-risk presumption, Article 55 duties and the Article 52 notification follow.

If You Stay Below the Line: You Are Still a Downstream Actor

Not becoming a model provider does not mean having no obligations. A company that fine-tunes lightly and deploys the result inside its product remains the provider of an AI system, with high-risk duties under Chapter III where Annex III applies, transparency duties under Article 50 from August 2, 2026, and AI literacy duties under Article 4. The original model provider's Annex XII documentation remains the foundation for that compliance, so fine-tuners should obtain it and keep records showing their modification stayed within the supplier's acceptable use terms and below the compute yardstick.

Practical Compute Accounting for Fine-Tuners

• Record the compute of every adaptation run — GPU hours, hardware type, and an estimated FLOPs figure with the estimation method noted.
• Where the original training compute is published or disclosed in the supplier's documentation, compare cumulative modification compute against the one-third yardstick and archive the comparison.
• Where it is not published, use the guidelines' proxy thresholds and document that you did.
• Aggregate honestly: successive fine-tuning campaigns on the same base accumulate toward the same question, and the analysis should be refreshed at each significant adaptation.

The Relationship with the Original Provider

Becoming the provider of a modified model does not erase the original provider's role: the upstream developer remains the provider of the original model and continues to owe its own Chapter V obligations for it. What changes is responsibility for the derivative. In practice the two providers are connected by documentation: the modifier's Annex XI file complements rather than reproduces the original one, and the modifier's Annex XII package for its own downstream customers will typically incorporate or reference the original provider's information. The Commission's guidelines encourage exactly this layered approach, and well-drafted model licences increasingly include cooperation clauses — access to documentation updates, notice of newly discovered limitations, and contact points for incident information — that make the layering workable.

Timing also follows the modification, not the base model. A new model created by significant modification is placed on the market when the modifier releases it, so the modifier cannot borrow the base model's pre-August-2025 transition status under Article 111(3); its obligations attach at its own release date. Fine-tuners adapting older bases should not assume the 2027 grace period travels with the weights.

A Concrete Example

A medical software company fine-tunes an open-weight 70-billion-parameter model on clinical documentation. The original provider discloses training compute of about 8 times 10^24 FLOPs; the company's fine-tuning consumes roughly 2 times 10^21 FLOPs — orders of magnitude below one third of the original. It does not become a GPAI model provider. It remains, however, the provider of an AI system used in a medical context, so it assesses Annex III and sectoral rules, keeps the supplier's Annex XII package in its technical file, and discloses AI interaction to clinicians. By contrast, a well-funded laboratory that takes the same base model and performs continued pre-training on 4 times 10^24 FLOPs of new data crosses the one-third yardstick, becomes the provider of a new GPAI model, and publishes a training summary covering its continued pre-training corpus while complementing the original technical documentation.

Common Pitfalls

The first pitfall is assuming any fine-tuning creates full Chapter V liability — it does not, and overclaiming provider status creates unnecessary obligations. The second is the opposite: heavy continued pre-training treated as mere fine-tuning, with no compute records to defend the classification when the AI Office asks. The third is forgetting the system layer: teams negotiate model-level questions carefully while shipping an Annex III high-risk application with no Chapter III workstream. The fourth is licence drift — fine-tuning and redistributing a model in ways the original licence does not permit, which creates contractual and copyright exposure independent of the AI Act. The fifth is data amnesia: fine-tuning datasets assembled informally from internal sources, with provenance that cannot support either the copyright policy or the training summary if provider status is triggered later.

Action Plan

Before each significant adaptation project, run a four-question check. How much compute will the modification consume relative to the base model's training compute or the proxy thresholds? What does the base model's licence and acceptable use policy permit? What is the provenance of the adaptation data, and could it support a training summary if needed? And what does the resulting system do — does Annex III or Article 50 attach? Write the answers down. For the large majority of fine-tuners the conclusion will be that they remain downstream actors with system-level duties only; the value of the exercise is being able to prove it.

Organisations running many adaptation projects should standardise the check as a lightweight gate in their MLOps pipeline: a one-page classification record per project, reviewed by whoever owns AI Act compliance, with compute figures pulled automatically from training infrastructure. The cost is minutes per project; the benefit is a defensible, consistent answer to the first question any supervisor or enterprise customer will ask — are you the provider of this model, or not?

The yardstick may also move. The Commission has indicated that the guidelines will be revisited as modification techniques evolve — parameter-efficient methods, merging, and distillation complicate any compute-based rule — so fine-tuning teams should re-check the current guidance at each major project rather than relying on a remembered number from 2025. A subscription to the AI Office's guidance updates, routed to whoever owns the classification gate, closes that loop at zero cost.

This article is for informational purposes only and does not constitute legal advice. Regulatory requirements change frequently — verify current rules with official sources. Built by Sawai Gyoseishoshi Office, Hiroshima, Japan.