The rules in force#
The obligations for providers of general-purpose AI (GPAI) models in the AI Act (Regulation (EU) 2024/1689), including technical documentation, information for downstream providers, a copyright policy and a public summary of training content under Article 53, have applied since 2 August 2025. On 18 July 2025 the Commission published guidelines on the scope of those obligations, including how they apply to organizations that modify an existing model.
Separately, the obligations for high-risk AI systems under Annex III — uses such as creditworthiness assessment or recruitment — were deferred by the Digital Omnibus on AI, adopted in June 2026, to 2 December 2027. Those obligations attach to the AI system and its use, regardless of whether you fine-tuned the model inside it.
This page summarizes public guidance as we understand it in October 2026 to help technical and risk teams plan. Confirm obligations with your counsel and compliance function.
The one-third compute rule#
According to the guidelines, a downstream modifier becomes the provider of the modified GPAI model only if the modification significantly changes the model’s generality, capabilities or systemic risk. As an indicative criterion, that is presumed when the training compute used for the modification exceeds one-third of the compute used to train the original model:
When the original compute is unknown, the guidelines substitute one-third of the presumption thresholds: FLOPs for GPAI models generally, or FLOPs for models presumed to carry systemic risk. Where a modifier does become a provider, its obligations are limited to the modification.
Where real fine-tunes fall#
Using our compute model ( for full fine-tuning plus activation recomputation and attention):
| Scenario | Training FLOPs | vs ⅓ × 10²³ | vs ⅓ of original pre-training |
|---|---|---|---|
| 7B model, full fine-tune, 60M tokens | 3.8e+18 | 8,879× below | 73,029× below |
| 72B model, full fine-tune, 300M tokens | 1.8e+20 | 188× below | — |
| Threshold (unknown original, non-systemic) | 3.3e+22 | — | — |
| Threshold (unknown original, systemic risk) | 3.3e+24 | — | — |
For the original compute of a 7B model we used the same rule of thumb: with parameters and pre-training tokens (the corpus size reported for Qwen2.5) gives FLOPs, so one-third is . A 60M-token fine-tune at is about 73,029 times smaller.
Even generous enterprise fine-tunes — hundreds of millions of tokens, several RL stages — remain orders of magnitude short of the threshold. Continued pre-training on billions of tokens is the scenario to check carefully.
What still applies to a fine-tuner#
- Your obligations as a deployer or provider of an AI system, which depend on the use case: transparency duties under Article 50 (applicable from August 2026), and high-risk obligations for Annex III uses from December 2027.
- Documentation you will need anyway. Downstream providers rely on the base-model provider’s documentation; keep it, alongside your own data card, model card and compute record.
- Licence terms of the base model, which are separate from the Act and sometimes stricter.
We record training compute (FLOPs, hardware, dates) for every engagement by default, so the threshold question can be answered with evidence rather than estimates.
References#
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Official Journal of the European Union.
- European Commission (2025). Guidelines on the scope of obligations for providers of general-purpose AI models under the AI Act, 18 July 2025. Summary: artificialintelligenceact.eu/gpai.
- Qwen Team (2024). Qwen2.5 Technical Report. arXiv:2412.15115.