---
title: The EU's Open AI Model Can Be Trained at Nearly a Trillion Parameters on a Public Supercomputer
description: Domyn trained a 951.9 billion parameter test version of EUROPA, the EU's open AI model, on JUPITER at 14 times its starting speed.
author: Darie Nani (Editor-in-Chief)
date: 2026-11-17T11:54:35.686Z
updated: 2026-11-17T12:29:05.330Z
canonical: https://www.sovereignmagazine.com/article/eu-open-ai-model-europa-domyn
image: https://cdn.nanimediahouse.com/jupiter-supercomputer-julich-racks.webp
categories: Artificial Intelligence
content_type: News
region: Europe, Germany
publication: Sovereign Magazine
about:
  - type: Organization
    name: Domyn
    description: Domyn is a Milan-based AI company, founded in 2016 as iGenius by Uljan Sharka. It builds AI models and software for finance, government and heavy industry, and leads the EUROPA consortium building the EU's open frontier AI model.
    url: https://www.domyn.com
    industry: Artificial Intelligence
---

Domyn, the Milan company the European Commission picked in June to build Europe's open frontier AI model, has shown that a model of nearly a trillion parameters can be trained on JUPITER, the publicly funded supercomputer in Jülich, Germany.

In a technical post published on 12 November, the company says it raised training speed 14-fold, from 90 to about 1,300 tokens per second on each GPU, by reworking its training setup for a machine designed for scientific simulation. At that rate, Domyn says, the model could be pre-trained in about four months. Domyn describes the result as a working setup ahead of the full training run, and no model has been released.

Domyn says the result shows that Europe can compete in frontier AI on supercomputers it has already paid for, while dedicated AI factories are still being built. Bids for [the EU's planned AI gigafactories](https://www.sovereignmagazine.com/article/eu-tech-sovereignty-five-months-on) are now due on 3 December for one of the two lots and on 4 February 2027 for the other.

## Domyn's Consortium Won the EU's Frontier AI Grand Challenge in June

The Commission launched its Frontier AI Grand Challenge in February 2026. It offered to fund a single project to train a model with more than 400 billion parameters, a scale it associates with the world's most advanced AI systems. On 19 June it chose EUROPA, a consortium led by Domyn with the German research institutes Fraunhofer IAIS and Fraunhofer IIS.

The prize is computing time: up to 2.5% of all the computing capacity of EuroHPC, the EU's network of public supercomputers, for one year. The winning bid promised an open-source model that works in all 24 official languages of the EU, and the rules gave extra points for both. Domyn says it will release the model's weights together with its training recipes.

> "Europe can lead in advanced AI on its own terms."
> — Henna Virkkunen, European Commission

Henna Virkkunen, the Commission's Executive Vice-President for Tech Sovereignty, Security and Democracy, said when the winner was announced that EUROPA would show "that we can match the best while staying true to our values."

Domyn, formerly called iGenius, was founded in Milan in 2016 by Uljan Sharka, its chief executive. It sells AI models to finance companies, governments and heavy industry, and Domyn says its backers include Abu Dhabi's G42, Eurizon Capital, Rabobank and BNY. Sharka told Reuters in June that the model would be released within a year, fully open source and reproducible, so that companies and governments could run it on their own computers at no cost.

## EUROPA's Test Version Has 951.9 Billion Parameters and 384 Experts per Layer

The version Domyn is testing has 951.9 billion parameters, more than twice the Commission's minimum. It is a mixture-of-experts model: each of its 74 layers has 384 smaller expert networks, and for each word or part of a word it processes, a router picks eight of them. Only about 32.8 billion parameters are used for each token, so the model needs much less computing per token than one of the same size that uses all its parameters every time.

Domyn describes these figures as the configuration it uses for experiments and says some values may change before the training run starts. It has not given a final size, beyond a target of about one trillion parameters. AI-BOOST described the winning proposal in June as a model of 500 billion parameters.

## Domyn Raised Training Speed on JUPITER From 90 to 1,293 Tokens per Second per GPU

JUPITER, at Forschungszentrum Jülich, is Europe's first supercomputer in the exascale class and [began operating in 2025](https://www.sovereignmagazine.com/article/europe-sovereign-ai-talk-vs-action). Its main section has 5,884 computing nodes, each with four Nvidia GH200 chips. Domyn says hundreds of other projects share it, and it has the same user access as any of them: it cannot change the operating system, the network or the way the nodes' memory is set up.

Domyn says models of this size have so far been trained mostly on dedicated clusters of Nvidia's newer Blackwell chips, which have more memory per chip and faster links between them. On JUPITER, data moves between nodes six times more slowly than between chips inside the same node. In Domyn's first configuration, one measured training step took 83.6 seconds, of which the chips spent 14.3 seconds computing and 68.3 seconds waiting for data to arrive.

Most of Domyn's changes reduced that waiting time. The team recomputed some intermediate results instead of storing them, which let each chip handle longer sequences of text, and moved stored data into the memory of the chips' built-in processors. It sent the model's expert weights between nodes in a smaller number format, and ran the heaviest calculations of the training software's optimiser on the graphics chips instead of the processors. The last change limited how much of each chip the network software could use, which raised the speed from 1,210 to 1,293 tokens per second.

**Chart: Domyn made training on JUPITER 14 times faster**

Training speed after selected changes, in tokens per second on each GPU. Higher is better.

| Change | Tokens per second per GPU |
| --- | --- |
| Starting setup | 90 |
| Recompute, longer sequences | 370 |
| Data moved to CPU memory | 508 |
| Optimiser maths on the GPU | 597 |
| Expert weights sent in FP8 | 898 |
| Smaller gradients | 1,022 |
| Memory fixes | 1,210 |
| Network software capped | 1,293 |

Median of training steps 4 to 10, except 370, which comes from one step of a run that stopped early. Intermediate changes are not shown.

Source: Domyn, technical blog post, 12 November 2026 (https://www.domyn.com/blog/training-1t-moe-eu-public-infrastructure-pytorch)

At about 1,300 tokens per second, one copy of the model running on 384 chips processes about 43 billion tokens a day. Domyn says it can run four copies side by side on 1,536 chips, and at the 89% efficiency it measured, that would mean about 150 billion tokens a day. Domyn says its fixes are already in PyTorch and TorchTitan, the open-source software it trains with. The TorchTitan changes have been merged, and two of its PyTorch changes were still under review on 17 November.

## The Winning Proposal Promised an Apache 2.0 Licence and Open Data Only

AI-BOOST, the EU-funded project that ran the challenge with the Commission and EuroHPC, says EUROPA will be trained only on open data and released under the Apache 2.0 licence, which lets anyone use, change and sell the model. Domyn has not yet published a list of the datasets it will use. In June, Sharka said Domyn would gather data from institutional partners and expected its first data agreements with governments within weeks.

How good the model will be is not known until training is finished and the model is tested. Domyn says early runs suggest one of its speed-ups, which forces an even split of work across the chips, can hurt training, and it has not yet decided whether to keep it. It also lists questions it has not yet answered, including whether training could go faster still with other settings. Its evidence that the smaller number formats do not harm training comes from comparison runs of ten steps, in which the results stayed within 0.01 of each other. Domyn says it will report on these questions as the training run goes on.

Europe already has one open model of this size: [Large 4](https://www.sovereignmagazine.com/article/mistral-large-4-open-weight-benchmarks), whose weights the Paris company Mistral made public on 29 October. EUROPA is meant to differ in two ways: it is paid for with public computing time, and its training recipes are to be published along with the model. On the timeline Sharka gave in June, it would be released by mid-2027.

Find out more at [domyn.com](https://www.domyn.com).

## FAQ

**Q: What does Domyn do?**
Domyn is a Milan-based AI company, founded in 2016 as iGenius by Uljan Sharka. It builds AI models and software for regulated sectors such as finance, government and heavy industry. In June 2026 it was chosen by the European Commission to lead EUROPA, the consortium building the EU's open frontier AI model with Fraunhofer IAIS and Fraunhofer IIS.

**Q: What is the JUPITER supercomputer?**
JUPITER is Europe's first supercomputer in the exascale class, run by the Jülich Supercomputing Centre in Germany as part of the EU's EuroHPC network. Its main section has 5,884 nodes, each with four Nvidia GH200 chips. Domyn says hundreds of other projects share it, and Domyn is using it for the trial runs of EUROPA.

**Q: Is there a European AI model?**
Yes. The Paris company Mistral made the weights of Large 4, a model of about a trillion parameters, public in late October 2026, and smaller open models such as EuroLLM and Switzerland's Apertus are also available. EUROPA, the model Domyn is building for the European Commission, is not yet trained; its chief executive said in June 2026 that it would be released within a year.
