EU Chooses Italian Startup Domyn to Train Open AI Model with Over 400 Billion Parameters in 24 Languages

A Bet on Computational Sovereignty

On Friday, June 19, the European Commission chose the EUROPA consortium, led by the Italian startup Domyn, as the winner of the Frontier AI Grand Challenge. The project will receive up to 2.5% of the total computing capacity of EuroHPC supercomputers for one year to train an open-source language model with over 400 billion parameters, developed in the 24 official languages of the European Union.

The initiative was launched in February 2026 by the European Commission in conjunction with EuroHPC, the European supercomputing network, with the stated aim of creating an AI model comparable in scale to the systems of OpenAI, Google, and Anthropic, but under European governance and with source code open to the public. The minimum scale required, 400 billion parameters, places the winning proposal in the same tier as GPT-4o and Llama 4 models.

The Consortium and the Model

The Frontier AI Grand Challenge required efficient architectures capable of operating at the frontier scale. The EUROPA model will utilize Mixture-of-Experts (MoE), a design that allows for parameter scaling without proportional growth in inference cost, the same approach adopted by Meta's Llama 4 family. At the end of the training, the model will be made available as open-source, differentiating it from the closed models of major American labs.

Domyn, based in Milan, leads the consortium. The company competes in a restricted segment of European frontier AI startups, alongside names like Paris-based Mistral AI and Germany's Aleph Alpha. The full composition of the consortium, including research partners and computing infrastructure, has not yet been disclosed by the European Commission.

Why an Italian Startup Leads

The Commission chose a consortium headed by Domyn rather than a university or public institute like the French INRIA or the German DFKI, signaling its desire to build native industrial capability, not just scientific expertise. This is the first time the European Commission directly finances the training of a frontier AI model instead of merely regulating the models produced by others. The distinction matters: regulating is about setting limits on what others do; training is about claiming the right to do so first.

EuroHPC boasts supercomputers among the top 10 globally. LUMI in Finland, Leonardo in Italy, and MareNostrum 5 in Spain are among the systems that could be allocated to the project. Reserving 2.5% of the collective capacity for one year represents the largest commitment of computing infrastructure ever made in the context of European AI policy.

Sovereignty vs. Performance: The Real Dilemma

The decision places Europe in strategic confrontation with the business models of major American labs. OpenAI, Anthropic, and Google offer cutting-edge models under proprietary licenses and rely on American infrastructure. Meta is the notable exception: Llama 4, open-source and in the same parameter range, serves as a direct technical reference for the EUROPA project.

The difference is not merely technical; it is one of governance. An EU-trained model on EuroHPC under an open license addresses institutional clients who currently cannot use GPT-4o in sensitive applications: public hospitals in Germany, courts in Portugal, security agencies in Poland. None of these sectors can outsource data to an American model without violating GDPR or national sovereignty restrictions. There are hundreds of public institutions across the 27 member countries that need language models for tasks ranging from medical screening to legal analysis, and currently, there is no local alternative in the frontier tier.

The legitimate skepticism is technical, not political: can an Italian startup leading a consortium with 2.5% of a shared cluster achieve parity with labs that have invested tens of billions in proprietary infrastructure and exclusive datasets? Mistral AI, a well-capitalized European company, has not yet launched a model in the 400 billion parameters tier. Domyn will have to accomplish what even the leading European AI entity has not.

If the EUROPA model achieves parity in at least twelve of the 24 languages on standard benchmarks like MMLU and MT-Bench by launch, the bet will appear visionary in retrospect. If it falls five points short of the leaders, the debate about contracting American models with contractual guarantees of sovereignty will resurface with renewed vigor. The technical outcome, expected by 2027, will define the terms of the sovereign AI dispute in the block for the next five years.