Fine-Tuning an Open-Weight Model Under the GPAI Rules
Alicia Labs, a Spanish startup, fine-tunes an open-weight 70B model on 2.1 trillion tokens of licensed Spanish legal text and releases the weights as 'Lex-70B' under its own name. It estimates 8×10^24 FLOPs for the fine-tune; the base model provider reported 3×10^25 FLOPs. Alicia publishes a model card, a training-data summary, and a copyright policy, but no red-team results or systemic-risk evaluation. Lex-70B has 40,000 EU downloads and one bank uses it for contract review. The Commission has not designated the base model. Alicia's counsel argues the company is a 'downstream modifier', not a provider, and that the 10^25 FLOP threshold is not met by its own compute.
1. Is Alicia a provider of a general-purpose AI model under the EU AI Act? Explain the triggers.
Read the worked answer
Alicia is likely a provider. The Act defines a provider of a general-purpose AI model as a person that develops a model or has one developed and places it on the market or puts it into service under its own name or trademark, for payment or free of charge. Alicia releases the weights under its own name, so the 'downstream modifier' label does not create a safe harbour. The relevant question is not only who pretrained the base model but who takes responsibility for the model as placed on the market. Fine-tuning can also amount to a substantial modification if it materially changes capabilities or behaviour, and releasing the result under a new name strengthens that conclusion. Alicia can rely on the base model provider's documentation only where it remains accurate and sufficient for Lex-70B; it cannot simply pass through the base model's paperwork. If Alicia merely used the model internally without placing it on the market, the analysis would differ, but public weight release and 40,000 EU downloads are placing on the market.
2. Does Lex-70B have systemic risk? Address the 10^25 FLOP presumption and how Alicia should reason about the base model's compute.
Read the worked answer
On Alicia's own fine-tuning compute of 8×10^24 FLOPs, the Article 51 presumption is not triggered, because the threshold is greater than 10^25 FLOPs. But that is not the end of the analysis. The provider must assess high-impact capabilities against the Annex XIII criteria, and the Commission can designate a model on that basis. The base model's 3×10^25 FLOPs is relevant because Lex-70B is derived from it. Alicia should document whether the fine-tune is a new model or a modification, and whether cumulative training compute includes the base pretraining. A conservative reading treats the base compute as part of the lineage and therefore treats the presumption as potentially met; at minimum, Alicia cannot assert '8×10^24 so no systemic risk' without addressing the base. If the base model had already been designated, Alicia would need to treat the fine-tune as likely in scope until it can rebut. If Alicia concludes it has systemic risk, it must notify the Commission without delay and comply with the additional evaluation, adversarial testing, incident reporting, and cybersecurity obligations.
3. What obligations and gaps should Alicia address before and after release, and does the bank's contract-review use change the model's classification?
Read the worked answer
As a GPAI provider, Alicia needs technical documentation, information for downstream providers, a copyright policy, and a sufficiently detailed summary of training content. If systemic risk applies, it also needs model evaluation, adversarial testing, serious-incident reporting, and cybersecurity protections. The gaps are significant: no red-team results, no systemic-risk evaluation, and no incident process. The model card and data summary are a start but do not substitute for the technical documentation and downstream information required by the Act. The bank's use for contract review does not turn Lex-70B into a high-risk AI system by itself, because contract review is not an Annex III use. High-risk classification attaches to the AI system and its intended purpose, not to the model in the abstract. If the bank integrates Lex-70B into a credit-scoring or employment system, that system may be high-risk and the bank may become a provider or deployer with separate obligations. Alicia should give the bank accurate instructions and warnings about limitations, and should monitor for foreseeable high-risk downstream uses rather than assume the bank's compliance is entirely the bank's problem.