Europae Laeva

One Europe, for the many.

European supercomputer JUPITER tests near-trillion-parameter training for open AI model

Domyn says a demonstration run on Europe's first exascale machine could pre-train the open EUROPA model in about four months, though only a fraction of its parameters are active at any moment.

Overhead view of a large server hall with rows of cabinets, glowing red lights and small engineers working among cables
Illustration generated with AI.

Luca Antiga of Domyn announced that JUPITER, a European public supercomputer, has been set up to train a model with about 1 trillion parameters. Antiga announced the setup on 12 November. In June the Commission chose EUROPA, a consortium led by the Italian company Domyn with Fraunhofer, as winner of its Frontier AI Grand Challenge.

In February 2026 the Commission and the EuroHPC Joint Undertaking launched the Frontier AI Grand Challenge, asking for one team able to train a model with at least 400 billion parameters. For one year, EUROPA gets up to 2.5% of EuroHPC's total computing capacity on one or more of its supercomputers. Antiga said that until then it had been unclear whether the machine could train such a model at all.

Domyn's first large-scale training trials for the open EUROPA model used a test configuration of 951.9 billion parameters, more than double the 400 billion minimum the Commission asked for. The model uses a mixture-of-experts design with 384 specialised sub-networks, calling on only eight of them for any given chunk of text. So just 32.8 billion of the nearly 952 billion parameters work at any moment.

The demonstration ran on JUPITER at the Jülich research centre in Germany, Europe's first exascale supercomputer, which is shared with hundreds of scientific projects. Domyn had to train on the machine as it was, without reconfiguring the hardware. Domyn's engineers rebuilt the training stack around PyTorch and TorchTitan for the run, and contributed fixes back to both projects. According to the company, EUROPA's pre-training in the demonstration setup is possible in about four months.

The consortium says the model and its training recipes will be released when EUROPA is published. The configuration it tested is a feasibility experiment, and some values may change before training starts. Domyn, for its part, says Europe needs more ambition, more data centres, more capital and more time. The gap is wide: according to Epoch AI, the United States received about 72% of the AI computing capacity installed worldwide between 2022 and June 2026, the EU about 5%. The Commission's answer so far is InvestAI, which Ursula von der Leyen presented in February 2025 to mobilise €200 billion, including €20 billion for AI gigafactories.