Fine-Tuning a Foundation Model Under the EU AI Act: When You Become a GPAI Provider (and the One-Third Compute Test)
Fine-Tuning a Foundation Model Under the EU AI Act: When You Become a GPAI Provider
A team downloads an open-weight model, fine-tunes it on company data and ships it in a product. Are they a deployer of someone else's model, or have they become a provider with their own obligations? For general-purpose AI (GPAI) models the Commission's guidelines give a surprisingly specific answer, and since 2 August 2026 the AI Office can enforce it.
First: is the model a GPAI model at all?
The Commission's guidelines presume a model is general-purpose if it was trained with more than 10^23 FLOP of compute and can generate language, text-to-image or text-to-video output. This is a presumption, not a bright line[1]: some models above it may not qualify, and some below it may still show sufficient generality. Models trained with more than 10^25 FLOP are presumed to pose systemic risk[2], which brings the heavier duties covered in our Safety and Security chapter guide.
The downstream modifier test
Modifying a GPAI model does not automatically make you its provider. According to the guidelines, you become one when the modification significantly changes the model's generality, capabilities or risk profile[1]. The guidelines add an indicative criterion: the compute used for the modification exceeds one third of the compute used to train the original model, or, if that is unknown, 3.33 x 10^22 FLOP[1].
| Scenario | Likely status | Why |
|---|---|---|
| Prompting or RAG on top of a hosted model | Deployer / downstream system builder | No change to the model itself |
| LoRA or adapter fine-tune on a small dataset | Usually still deployer | Compute far below one-third threshold |
| Large continued pre-training near or above one-third of original compute | Likely modifier-provider | Meets the indicative criterion |
| Fine-tune that changes risk profile materially, such as removing safeguards | Case-by-case | Criterion is indicative, not exclusive |
The one-third figure is an indicator, not a safe harbour. Document why a lower-compute modification does not significantly change generality, capabilities or risk.
What a modifier-provider owes
If you cross the line, you take on the original provider's compliance role, but only for your modification[1]. The core Article 53 duties are:
- technical documentation of the model,
- information and documentation for downstream providers,
- a copyright compliance policy, and
- a public summary of training content.
The open-source exemption covers only the first two of these[1], so the copyright policy and training summary remain. A modifier established outside the EU must also appoint an authorised representative in the EU[1]. For details on the copyright duties, see our copyright chapter guide, and for how far open source gets you, our Article 2(12) guide.
Find the original model's training compute from its model card or documentation. If unavailable, use the 3.33 x 10^22 FLOP fallback.
Record compute for fine-tuning, continued pre-training and any reinforcement stages, with dates and hardware.
Compare benchmarks and red-team results before and after. Note any change in refusal behaviour or dangerous capability.
Write a short determination memo: provider or not, with the reasoning. Review it whenever you retrain.
Prepare documentation, downstream information, copyright policy and training summary for the modified model, and appoint an EU representative if needed.
Why this matters more now
Several dates frame the risk:
- GPAI obligations have applied since 2 August 2025[1].
- Since 2 August 2026 the Commission can enforce, with fines of up to EUR 15 million or 3% of worldwide annual turnover[3][4].
- Models already on the market before 2 August 2025 have until 2 August 2027[1][4] to comply.
The AI Office's powers include formal documentation requests, model evaluations through independent experts, restricting a model's availability and opening investigations without a complaint[3]. Reports so far indicate the Office has not named a first target, but a modifier with no documentation is an easy one to ask.
The Digital Omnibus also gives the AI Office exclusive competence over AI systems built on a GPAI model where model and system come from the same provider[4]. If you both fine-tune and ship the product, your supervisor is likely to be Brussels rather than a national authority.
Practical advice
- Keep a modification log for every fine-tune: base model, data sources, compute, evaluation deltas.
- Ask upstream providers for their Article 53 documentation, because as a downstream builder you rely on it.
- Check whether the base model's provider signed the GPAI Code of Practice, which gives a defined route to showing compliance.
- Treat licence terms and EU-law duties separately; an open licence does not remove Article 53 obligations.
Does fine-tuning with LoRA make me a GPAI provider?
Usually not. Such fine-tunes tend to use far less than one third of the original compute and do not significantly change generality, though you should document that conclusion.
What if the original compute is unknown?
The guidelines use 3.33 x 10^22 FLOP as the fallback benchmark for the modification.
Do I need to redo all provider documentation?
No. A modifier-provider is responsible only for the modification, not the whole model.
Does open-source release avoid the obligations?
Only partly. The exemption covers technical documentation and downstream information, but not the copyright policy and training content summary.
Does this apply to non-EU companies?
Yes, if the model is placed on the EU market. A non-EU modifier-provider must appoint an authorised representative in the EU.
This article is general information, not legal advice.
Sources: Arnold & Porter on GPAI provider obligations; Pondero on GPAI enforcement from 2 August 2026; European Commission GPAI guidelines; ActuIA on Omnibus 2026/1744.
Related reading
Article 14 EU AI Act: Engineering Human Oversight That Actually Works
Article 14 is a design obligation, not an HR policy. Here's how to engineer the five oversight capabilities, split provider vs deployer duties, handle the biometric two-person rule, and build an audit-ready evidence trail.
EU AI Act Article 13: What Your Instructions for Use Must Contain, and What Deployers Should Demand From Vendors
Article 13 turns the instructions for use into the contract between a high-risk AI provider and its deployers. Here is what Article 13(3) requires, the evidence deployers should ask for, and where harmonised standards stand after 14 October 2026.
GPAI Models With Systemic Risk: The Safety and Security Chapter of the Code of Practice, Explained
Providers of the most capable general-purpose AI models carry extra duties. Here is what the Safety and Security chapter of the GPAI Code of Practice asks for, the incident timelines, and the dates that matter now the AI Office can enforce.