The European Commission has put one of its biggest AI bets on Domyn, a Milan company that most people outside European enterprise software knew until last year as iGenius. On June 19, the Commission named the Domyn-led EUROPA consortium as the winner of its Frontier AI Grand Challenge. The prize is the right, and the compute, to train an open-source model of more than 400 billion parameters covering all 24 official EU languages.
CEO Uljan Sharka told Reuters the model will ship within a year. Trained from scratch. On shared public supercomputers. Through a consortium that has to coordinate across institutions and governments.
About that timeline
A year is the number to watch here. The grand challenge only launched in February 2026, and a winner was picked roughly four months later, which is quick by the standards of EU procurement. You can read Sharka's twelve-month promise as confidence or as the only answer you can give once Brussels makes your company the face of European AI. Probably both.
Training a 400-billion-parameter multilingual model from a cold start in twelve months would stretch a far larger outfit. The spec required a Mixture-of-Experts architecture, and the consortium gets up to 2.5% of EuroHPC's total computing capacity for one year, which one writeup called the largest single compute allocation ever handed to a European AI project. That superlative is worth a pinch of salt; it traces to secondary commentary, not the Commission's own announcement.
Who's actually in this
Domyn isn't doing it alone. The consortium was set up with Germany's Fraunhofer research organisation, and the company already ships models into regulated sectors: a 263-billion-parameter model called Domyn-Large, plus earlier ones named Italia and Colosseum. So this is a commercial player with a track record, not a research lab assembling from nothing.
Sharka's pitch is that companies and governments will run the thing on their own hardware at no usage cost. True enough on licensing. But open weights and free operation are different animals, and standing up a 400-billion-parameter system reliably is its own expensive engineering problem. The license is free. The electricity bill is not.
Why now
The sovereignty framing isn't abstract. Over the past year a string of European governments moved against Chinese AI tools: Italy's regulator blocked DeepSeek from app stores in early 2025, and the Czech Republic's government banned DeepSeek products across state administration in July, citing the risk of Chinese state access to user data. Most leading US models, meanwhile, stay proprietary and remote, and US export controls have rattled European buyers who don't love depending on a stack built in San Francisco or Shenzhen.
That's the gap EUROPA is meant to fill: a model you can audit, host locally, and actually control. Whether it can match frontier performance while doing all that, on this timeline, is the open question. Executive Vice-President Henna Virkkunen says the point is to prove Europe can match the best on its own terms. Easy to say at a launch. Harder to ship.
The model does not exist yet. Domyn says first data-access agreements with European governments are expected in the coming weeks, with a release targeted within the year.
