The obligations on general-purpose models already apply: from when and for whom
The title of this article, with that "2027", suggests a deadline that is still far away. It is not: Chapter V of Regulation (EU) 2024/1689, the one on general-purpose AI (GPAI) models, applies from 2 August 2025 to models placed on the market from that day. Whoever already had a model on the market before 2 August 2025 has until 2 August 2027 to comply (Article 111(3)), and that is probably where the 2027 comes from. The specific fines for providers of these models (Article 101) are applicable from 2 August 2026 and are imposed by the Commission, not by the individual Member States. Regulation (EU) 2026/1744 moved the high-risk dates, not these.
The question for a developer is a different one: am I the "provider" of a GPAI model? Almost always no. Whoever calls an API or downloads open weights and uses them is, for this chapter, downstream. The obligations concern whoever develops the model and places it on the market, and whoever modifies it enough to become a provider in turn. Here you see how to tell the difference, what providers must produce, how the threshold is computed and where the text and the sources stop. The role of provider of an AI system, different from that of provider of a model, is in the first article; the previous topic, civil liability, is in the second.
What a general-purpose model is
Article 3 defines the GPAI model as an AI model, including where it is 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 and that can be integrated into a variety of downstream systems or applications; AI models used for research, development or prototyping activities before they are placed on the market are excluded. The definition contains no numbers, which is why the Commission published guidelines (Communication C(2025) 7719) with an indicative criterion: a model is a GPAI model if its training compute is greater than 1023 FLOP and it can generate language (whether in the form of text, which includes code, or audio), text-to-image or text-to-video. The guidelines explain that 1023 FLOP corresponds to the approximate amount of compute typically used to train a model with one billion parameters on a large amount of data, and they state that the criterion is indicative and may change.
According to Annex A.2.2 of the same guidelines, compute is estimated as C ≈ 6 · P · D for some dense large language models based on the transformer architecture, where P is the number of parameters and D the number of training tokens. The following program applies the formula and reproduces the numerical example of the guidelines (model A: 3.8 billion parameters and 3.3 trillion tokens, about 7.5 · 1022 FLOP).
"""Training compute estimate according to Annex A.2.2 of the Commission
guidelines C(2025) 7719: C ~ 6 * P * D, valid for dense transformer models."""
SOGLIA_INDICATIVA_GPAI = 1e23 # guidelines, paragraph 17 (with language generation)
SOGLIA_PRESUNZIONE = 1e25 # Art. 51(2), Regulation (EU) 2024/1689
def calcolo(parametri, token):
return 6 * parametri * token
def classifica(nome, parametri, token):
c = calcolo(parametri, token)
if c > SOGLIA_PRESUNZIONE:
esito = "presumed high-impact capabilities (Art. 51(2))"
elif c > SOGLIA_INDICATIVA_GPAI:
esito = "above the indicative GPAI criterion, below the systemic-risk presumption"
else:
esito = "below the indicative criterion (a hint, not proof)"
print(f"{nome:<34} C = {c:.2e} FLOP -> {esito}")
return c
def fine_tuning(nome, c_originale, c_modifica):
quota = c_modifica / c_originale
esito = "becomes provider of the modified model" if quota > 1 / 3 else "stays downstream: no GPAI provider obligations"
print(f"{nome:<34} modification/original = {quota:.4f} -> {esito}")
# Numerical example from the guidelines (model A: 3.8 billion parameters, 3.3 trillion tokens)
c_a = classifica("Model A (guidelines example)", 3.8e9, 3.3e12)
classifica("7-billion model, 2 T tokens", 7e9, 2e12)
classifica("400-billion model, 15 T tokens", 4e11, 1.5e13)
# A fine-tuning run with 50 billion tokens on model A
fine_tuning("Light fine-tuning (50 B tokens)", c_a, calcolo(3.8e9, 5e10))
# An extensive retraining with 2 T tokens
fine_tuning("Extensive retraining (2 T tokens)", c_a, calcolo(3.8e9, 2e12))
# Output (python3 -I, Python 3.14.4):
# Model A (guidelines example) C = 7.52e+22 FLOP -> below the indicative criterion (a hint, not proof)
# 7-billion model, 2 T tokens C = 8.40e+22 FLOP -> below the indicative criterion (a hint, not proof)
# 400-billion model, 15 T tokens C = 3.60e+25 FLOP -> presumed high-impact capabilities (Art. 51(2))
# Light fine-tuning (50 B tokens) modification/original = 0.0152 -> stays downstream: no GPAI provider obligations
# Extensive retraining (2 T tokens) modification/original = 0.6061 -> becomes provider of the modified model
The first two constants are the thresholds of the text: the indicative one of the guidelines and the one of Article 51(2), above which the model is presumed to have high-impact capabilities and therefore systemic risk. calcolo ("computation") is the formula; classifica ("classify") compares the result with the two thresholds. The output contains a detail that matters: model A of the guidelines sits below 1023 although it is one of the examples that informed the threshold, so "below the threshold" is not proof of not being GPAI but a hint, and that is how the program words it in the output. The formula also has a declared limit: the guidelines present it as an approximation valid for some dense large transformer models; for other cases they describe a second method, based on hardware (number of GPUs, training hours, peak performance and utilisation rate).
Who is the model provider: three scenarios
The provider is whoever develops the model, or has it developed, and places it on the Union market. This leaves out most of those who work with models.
- You call the API of someone else's model. You are not the provider of the model. You are a provider or deployer of an AI system, with the obligations of the first article, and you are entitled to receive from the model provider the information and documentation to understand its capabilities and limitations and to fulfil your own obligations (Article 53(1), point (b), and Annex XII).
- You run an open-weights model, without modifying it. Not even in this case are you the provider of the model. The original provider remains whoever released it; the exemptions of Article 53(2) are theirs.
- You modify a model with additional training (fine-tuning). You become the provider of the modified model only if the modification leads to a significant change in its generality, capabilities or systemic risk. The guidelines propose an indicative criterion: the training compute used for the modification is greater than a third of the training compute of the original model.
In the last case, if the original compute is neither known nor estimable, the threshold is a third of 1025 FLOP for models with systemic risk, otherwise a third of 1023 FLOP. The guidelines acknowledge that few modifications currently meet the threshold and that the criterion is primarily forward-looking. Whoever exceeds it has obligations limited to the modification: technical documentation complemented with the information on the modification, copyright policy and training-content summary limited to the data used for the modification, and the obligation to appoint an authorised representative if established outside the Union (Article 54).
In the program, fine_tuning divides the compute of the modification by that of the original and compares it with one third. A fine-tuning run with 50 billion tokens on model A is 1.5% of the original compute: it stays downstream. A retraining with 2 trillion tokens is 60%: it exceeds the threshold. The practical conclusion runs against what you often read: using fine-tuning "to avoid the obligations" is not a strategy, because the threshold depends on compute, not on the name of the operation; but it is equally true that the typical fine-tuning of an application team sits well below a third of the original compute.
What providers must do (Article 53)
All providers of GPAI models must meet four obligations, and the boundary with AI systems is sharp: they are documentation and intellectual-property obligations, not safety testing as for high-risk.
- Draw up and keep up to date the technical documentation of the model, containing at a minimum the information set out in Annex XI, in order to provide it, upon request, to the AI Office and the national competent authorities.
- Draw up, keep up to date and make available information and documentation for providers of AI systems that intend to integrate the model, containing at a minimum the elements set out in Annex XII.
- Put in place a policy to comply with Union law on copyright and related rights, and in particular to identify and comply with, including through state-of-the-art technologies, a reservation of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790.
- Draw up and make publicly available a sufficiently detailed summary about the content used for training, according to a template provided by the AI Office.
The first and the second are the two faces of the same supply chain: Annex XI is the sheet for the authorities (architecture and number of parameters, training data with type and provenance, computational resources, known or estimated energy consumption), Annex XII is the sheet for whoever integrates the model. The third and the fourth concern whoever produced the data. Article 53(2) exempts from points (a) and (b) providers of models released under a free and open-source licence that allows access, usage, modification and distribution and whose parameters, including the weights, the information on the model architecture and the information on model usage, are made publicly available; the exception does not apply to models with systemic risk, and it does not touch the copyright policy or the data summary.
Systemic risk, notification and codes of practice
A GPAI model has systemic risk if it has high-impact capabilities, evaluated on the basis of appropriate technical tools and methodologies, or by decision of the Commission. High-impact capabilities are presumed when the cumulative amount of computation used for training is greater than 1025 FLOP (Article 51(2)). The provider must notify the Commission without delay and in any event within two weeks after that requirement is met or it becomes known that it will be met (Article 52(1)). In addition, Article 55 requires model evaluation in accordance with standardised protocols, including conducting and documenting adversarial testing, the assessment and mitigation of systemic risks at Union level, the tracking, documenting and reporting without undue delay of serious incidents, and an adequate level of cybersecurity protection for the model and its physical infrastructure.
Providers may rely on codes of practice to demonstrate compliance until a harmonised standard is published; those who do not adhere to an approved code must demonstrate alternative adequate means of compliance (Article 53(4)). The GPAI Code of Practice, published on 10 July 2025, has three chapters: Transparency (with a Model Documentation Form), Copyright, and Safety and Security. The first two serve all providers for Article 53; the third concerns only the small number of providers of the most advanced models subject to Article 55. The Commission and the AI Board have confirmed that the code is an adequate voluntary tool. Adhering is not an obligation, but it shifts the burden: whoever adheres can demonstrate compliance with the code, whoever does not must argue for an alternative.
The fines: the Commission may impose on providers fines not exceeding 3% of annual total worldwide turnover or EUR 15 000 000, whichever is higher, when it finds that the provider intentionally or negligently infringed the provisions, failed to comply with a request for a document or for information, or supplied incorrect, incomplete or misleading information, and the supervision of Chapter V is the exclusive competence of the Commission, which entrusts it to the AI Office (Article 88). For an Italian company that only builds downstream applications, the reference authorities for the rest of the regulation are AgID (the Agency for Digital Italy) and ACN (the National Cybersecurity Agency), designated by Italian Law 132/2025 (Article 20; my summary of the Italian text); for the model, they are not.
Frequently asked questions
My model has 500 million parameters: is it still GPAI?
It depends on the training compute and on what it can generate, not on the number of parameters alone. With the 6PD formula, 500 million parameters exceed 1023 FLOP only with more than 3.3·1013 tokens (about 33 trillion), because 6 · 5·108 · D must be greater than 1023. The criterion is indicative: if you are near the threshold, keep the estimate and the method.
If I use a proprietary API, do I have to comply with the obligations of Chapter V?
No, those belong to the model provider. You must, however, use the information they must provide to you (Annex XII) to fulfil your own obligations as provider or deployer of a system.
Can I fine-tune to avoid the obligations?
Fine-tuning neither creates nor removes obligations by itself: if the compute of the modification exceeds a third of that of the original model (or the fallback threshold) you become the provider of the modified model, otherwise you do not. The criterion is indicative and may change.
Which authority applies these rules in Italy?
For Chapter V, the European Commission through the AI Office (Article 88). For the rest, Italian Law 132/2025 designates AgID and ACN as national AI authorities, with the Bank of Italy, CONSOB and IVASS for financial cases (my summary of the Italian text).
Is a third-party audit needed to be compliant?
Article 53 does not provide for a third-party audit; it asks for documentation, a copyright policy and a data summary. For models with systemic risk, Article 55 asks for documented evaluations and adversarial testing; whoever hands part of it to third parties does so by choice or because of a commitment made under the code, which I have not analysed in detail.
What you can do now
You know that the obligations on GPAI models apply from 2 August 2025 and not in 2027, and why that number circulates; you can estimate the compute of a training run with the formula of the guidelines and compare it with the two thresholds; you can say whether you are provider of the model, of a system or only deployer, and apply the one-third criterion to fine-tuning; and you know what providers must publish and document. The next article moves on to generated content: the marking under Article 50. For the practical use of models in a company there is also Enterprise LLMs: RAG, Fine-Tuning and AI Guardrails.
Sources
- Regulation (EU) 2024/1689, consolidated text of 27.7.2026 - Articles 3, 51-56, 88, 101, 111 and Annexes XI and XII, read in the English text, consulted on 2026-11-13
- European Commission, guidelines on the scope of obligations for providers of general-purpose AI models, C(2025) 7719 final - indicative criterion of 10^23 FLOP, one-third threshold for modifications, Annex A.2.2 on the 6PD formula, consulted on 2026-11-13
- European Commission, General-Purpose AI Code of Practice - publication of 10 July 2025, the three chapters, voluntary role, consulted on 2026-11-13
- European Commission, page on the guidelines for providers of GPAI models - application dates and enforcement powers from 2 August 2026, consulted on 2026-11-13
- Legge 23 settembre 2025, n. 132 (Italian Law No 132 of 23 September 2025) on Normattiva - Article 20 on the national authorities, Italian text only, my own summary, consulted on 2026-11-13
The listing was run with Python 3.14.4, with identical output in two runs. The detail of the individual commitments of the Code chapters and the list of signatories were not verified; the guidelines cited are version C(2025) 7719 read on 2026-11-13.

Commenti
Caricamento commenti...
Sign in to comment