- Lead. The European Commission selected the EUROPA consortium, led by Italian AI startup Domyn, as the winner of its Frontier AI Grand Challenge — a competition to build a 400-billion-parameter open-source model covering all 24 official EU languages.
- Fact. The winning consortium will receive up to 2.5 percent of the total computing capacity of EuroHPC AI-optimised supercomputers for one year to train the model, which will be freely available to businesses, researchers, and public bodies across the EU.
- Stake. The project is the most direct expression yet of the EU’s push for AI sovereignty — a capability gap the bloc has struggled to close as US and Chinese labs have pulled further ahead at the frontier.
The Frontier AI Grand Challenge was launched by the Commission in February 2026 under the Apply AI Strategy, inviting Europe’s leading AI innovators to propose models capable of operating at or near the global frontier — defined as a minimum of 400 billion parameters. The threshold matters: it places the target in the same class as the largest models currently deployed by US labs, rather than the smaller, domain-specific models that European institutions have traditionally funded.
Domyn and the EUROPA Model
Domyn, the Italian company leading the EUROPA consortium, beat out competing bids to win the designation, which was announced by the European Commission on June 19. The project’s defining commitment is linguistic coverage: the model will support all 24 of the EU’s official languages, a design choice that distinguishes it from frontier models built primarily on English-dominant training data.
The project will be openly available — not licensed under commercial terms or restricted to specific institutions — and is intended to become a shared infrastructure asset for European businesses, universities, public administrations, and research bodies. Henna Virkkunen, the Commission’s Executive Vice-President for Tech Sovereignty, Security and Democracy, framed the selection as a proof of concept: “Europe can lead in advanced AI on its own terms… showing that we can match the best while staying true to our values.”
The Compute Question
Training a 400-billion-parameter model requires substantial compute. The Commission’s answer is access to EuroHPC — the EU’s network of AI-optimised supercomputers — for up to 2.5 percent of total capacity for one year. That allocation is large by European standards, though the absolute scale still falls short of what the leading US labs devote to individual training runs. Whether it is sufficient to produce a genuinely frontier-class model in multilingual settings, rather than a capable but second-tier one, will only become clear once training begins.
The challenge reflects a broader pattern in EU AI policy: regulatory frameworks like the AI Act set the guardrails, while industrial initiatives like EUROPA try to ensure European actors have competitive models to deploy within them. Earlier EU orders forcing open AI distribution channels addressed access; EUROPA addresses supply. Together they sketch the outline of a two-sided industrial policy for a sector Europe has not yet managed to lead.