Fine-tuning needs a provider-role record

The dangerous fine-tuning decision is often made before the first training run. A team chooses a base model, prepares a specialist data set and books compute, but nobody records whether the resulting model could give the company a new legal role.

That omission matters now. The European Commission’s enforcement powers for providers of general-purpose AI models began to apply on 2 August 2026. The same week, the AI Act Service Desk clarified the position for downstream modifiers: fine-tuning does not automatically turn them into providers of a new general-purpose AI model, but a significant change can.

For most SMEs, the practical conclusion is reassuring but not passive. Ordinary adaptation is unlikely to cross the Commission’s indicative compute line. The company still needs a short record showing which model it changed, how much training it performed, what capabilities moved and how the result will be used or supplied.

Fine-tuning is not the legal test

The AI Act distinguishes the model from the AI system in which it is used. A general-purpose AI model is the trained component capable of serving a wide range of tasks. The system adds an interface, workflow, tools, permissions and an intended use.

Fine-tuning describes a technical operation. It does not by itself answer who is the provider. According to the Commission’s guidelines, a downstream modifier becomes the provider of the modified general-purpose AI model only when the modification leads to a significant change in the model’s generality, capabilities or systemic risk.

The Commission uses training compute as an indicative proxy. A modification using more than one-third of the original model’s training compute points towards a new provider role. That is deliberately a high threshold. The Commission says most fine-tuning, adaptations and minor changes will not meet it.

This is guidance, not a new article inserted into the Regulation. The Commission says the guidelines are not legally binding and that only the Court of Justice can give an authoritative interpretation. An SME should therefore treat the compute ratio as evidence within a role assessment, not as a magic exemption.

The denominator is part of the evidence

A ratio is useless without its denominator. Before tuning, ask the original provider for the training compute of the exact base model or a usable estimate. Keep the response, the model identifier and the date. A family name such as “Model X” is not enough if several sizes or versions exist.

The Service Desk also addresses the case where the downstream modifier cannot reasonably know or estimate the original value. It gives two replacement reference points:

• for a base model with systemic risk, compare the modification with one-third of 10^25 floating-point operations;
• for another general-purpose AI model, compare it with one-third of 10^23 floating-point operations.

These are approximately 3.33 × 10^24 and 3.33 × 10^22 FLOP respectively. They are fallback tests for the modifier question, not budgets an SME should aim to consume.

Question	Evidence to keep	Why it matters
What is the base model?	Provider, model name, version, weights or API identifier, licence	Joins the decision to one model rather than a product family.
Is it a GPAI model?	Provider documentation, capabilities and available training-compute information	The special Chapter V duties concern general-purpose AI models.
What changes?	Training method, data set classes, objectives, evaluations and release notes	The legal question is whether generality, capability or systemic risk changes significantly.
How much compute is used?	Planned and actual training FLOP, hardware-hours, method and assumptions	Supports the Commission’s indicative one-third comparison.
Where will it go?	Internal use, customer service, API, download, embedded product or public release	“Placing on the market” can occur through several technical routes.

Source: European Commission GPAI guidelines and AI Act Service Desk FAQs. Last verified 2026-08-08.

Do not reconstruct this record after deployment from a cloud invoice alone. Hardware-hours can help estimate compute, but the estimate also depends on the accelerator, utilisation, training method and number of operations counted. Record the method while the engineers still know it.

Internal does not always mean outside the market

The Commission’s Q&A gives a useful low-risk example: an employer buying licences for a model already available in the Union is neither the provider of that model nor required to verify the provider’s compliance merely because staff use it.

The boundary changes when a business develops or modifies something under its own name or control. The Commission interprets “placing on the market” broadly enough to include APIs, downloads, cloud access and integration into applications. It can also include internal use where the model is essential to a product or service supplied to third parties, or affects the rights of people in the Union.

This is why a deployment label cannot settle the role:

SME activity	Model-provider question	Separate system question
Staff use a licensed third-party model	Normally no new model-provider role from the licence alone.	The SME may still be a deployer of an AI system in its work.
Retrieval or workflow logic around an unchanged model	No model modification has occurred merely because context or tools were added.	The assembled product may be an AI system with provider or deployer duties.
Modest domain fine-tuning	Usually below the Commission’s high indicative threshold, but document the assessment.	Intended use, transparency and risk classification still need review.
Significant modification released under the SME’s control	Assess whether the SME has become provider of the modified GPAI model.	Assess the resulting AI system independently if one is supplied or used.

Source: European Commission GPAI Q&A and Regulation (EU) 2024/1689. Last verified 2026-08-08.

The final column is the one teams most often lose. The Commission states explicitly that even when a downstream entity is not the provider of the general-purpose AI model, it must still comply with the relevant requirements for the AI system it builds around that model.

Keep two files

The cleanest operating pattern is to separate the model modification file from the system file.

The model modification file answers a narrow set of questions:

1. Who supplied the base model and under what version and licence?
2. What training or modification did the SME perform?
3. Which new data sources entered that work?
4. What compute was planned and actually consumed?
5. Did generality, capability or systemic risk change significantly?
6. Is the result made available under the SME’s name or control?
7. What provider-role conclusion was approved, by whom and on what date?

The system file follows the application: its intended purpose, users, tools, data access, human oversight, logs, transparency message, prohibited uses and risk classification. One file can remain small while the other grows. Combining them tends to obscure both decisions.

Parameter-efficient fine-tuning does not deserve a one-line dismissal simply because it changes relatively few weights. It may use little compute and remain far below the indicative threshold, yet still alter behaviour in a business-critical domain. That behavioural change belongs in the system’s testing and release controls even when it does not create a new GPAI-provider role.

Crossing the line does not recreate the base model

If a downstream modifier does become the provider of a modified general-purpose AI model, the AI Act does not require it to recreate every document belonging to the original training run. Recital 109 says the obligations should be limited to the modification or fine-tuning.

In practice, the new provider should expect to document the changes and new training data, keep technical information for authorities, supply useful information to downstream system providers, maintain a Union copyright policy and publish the required summary of training content for its modification. A provider established outside the EU may also need an authorised representative before placing the model on the Union market.

The original provider’s documentation therefore becomes an input, not a substitute. Procurement should request at least:

• an exact model and version identifier;
• the intended tasks, known limits and acceptable-use terms;
• technical integration and input-output specifications;
• the original provider’s training-content summary and copyright position;
• available training-compute information or a reasoned basis for estimating it;
• change notices that let the SME tell when the base has moved.

If the supplier cannot provide the denominator, record that fact before using the Service Desk fallback. If it cannot identify the model version or licensing route, the problem is more basic than the one-third calculation.

Sign the role decision before release

A provider-role record can fit on two pages. It should name the legal entity, base model, modification, compute estimate, capability evidence, intended distribution and separate system role. Attach the technical calculation rather than copying every training log into the decision.

Have engineering own the compute and evaluation facts. Have the product owner describe the intended release and customers. Have whoever owns legal or compliance review the role conclusion. Set a trigger for reassessment when the base model, tuning method, data volume, capability envelope or distribution route changes.

Then place the signed record beside the model artefact and release note. The useful question is not “did we fine-tune?” It is “what changed, how do we know, and which role follows from that evidence?” Answer it before the modified model acquires its first customer.