Europe’s bid to build a home-grown artificial intelligence model has reached an important milestone. Domyn, the Italian company chosen by the European Commission to develop an open EU model called EUROPA, has published the technical details of its first large-scale training trials and the scale is striking. The test setup runs at nearly a trillion internal settings, more than double the minimum the Commission asked for.
What EUROPA Is Meant to Be
The project grew out of the Frontier AI Grand Challenge, a contest the Commission launched in February 2026 together with the EuroHPC Joint Undertaking, the body that runs Europe’s network of public supercomputers. The brief was to find one team capable of training a “frontier” model (a top-tier AI system) with at least 400 billion parameters. The prize: up to 2.5% of EuroHPC’s total computing power for a year.
In June, the Commission named its winner, the EUROPA consortium, led by the Milan firm Domyn (formerly iGenius) alongside two German institutes from the Fraunhofer research network. The goal is to create an open model that governments, researchers and businesses across the bloc can download and run themselves, and one that works across all 24 official EU languages. “Europe can lead in advanced AI on its own terms,” said Henna Virkkunen, the Commission’s executive vice-president for tech sovereignty.
Making Sense of the Numbers
Parameters are the internal settings a model adjusts as it learns; roughly speaking, more of them means more capacity. Domyn’s test configuration has 951.9 billion, putting EUROPA in the same bracket as today’s leading commercial systems.
The model uses a design called “mixture of experts.” Rather than firing up the entire network for every piece of text, it splits into 384 specialised sub-networks and calls on only eight of them for any given chunk of text. That means just 32.8 billion of the nearly 952 billion parameters are actually working at any moment, a trick that keeps the computing demands manageable.
Wringing Speed Out of a Shared Machine
The more interesting story may be where EUROPA is being trained. Models this size are usually built on dedicated clusters of the newest chips. Domyn instead used JUPITER, Europe’s first exascale supercomputer, housed at the Jülich research centre in Germany and shared with hundreds of scientific projects. On a machine built for simulations rather than AI, and with only ordinary user access, Domyn could not reconfigure the hardware to suit its needs.
The main obstacle was the network linking the machine’s thousands of nodes, which is far slower than the connections inside a single node. At first, each graphics processor managed just 90 tokens (units of text) per second, idling while it waited for data. Through a series of engineering fixes, including compressing the data moving across the network and shifting heavy calculations between processors, Domyn lifted that figure to 1,293 tokens per second, a fourteen-fold improvement. The company fed its improvements back into PyTorch, the open-source software widely used to train AI. On that basis, the Italian daily Quotidiano Nazionale estimated a full training run would take around four months.
A Demonstration, Not a Finished Model
Important caveats remain. For all the impressive numbers, this is still an engineering demonstration rather than a completed model. The measurements come from short trial runs, and Domyn admits some settings may change before full training begins. The company has said nothing yet about the training data, the balance between languages, expected performance, or the licence under which it will be released. The weights, the files that would let anyone run EUROPA, have not been published.
Europe is not putting all its eggs in one basket. At least three “sovereign” AI projects are running in parallel, and in the private sector French firm Mistral released a roughly trillion-parameter model of its own in October. The pressure is real: Mistral’s chief executive, Arthur Mensch, noted this month that only a fifth of European digital services run on European infrastructure. EUROPA is a bet that Europe can change that, but the model, as one observer put it, has yet to write its first line of weights.
European Commission (2026). Commission selects EUROPA consortium as the winner of the Frontier AI Grand Challenge. 19 June 2026. https://digital-strategy.ec.europa.eu/en/news/commission-selects-europa-consortium-winner-frontier-ai-grand-challenge-project-build-european-open