A large language model that treats Maltese as seriously as English: that is the ambition proclaimed by the European Union, which is now pursuing three separate projects. The Commission published its own institutional model in July 2026, designated a consortium in June to build a "frontier" model in the 24 official languages, and has been funding the OpenEuroLLM research project since 2025. Three different approaches, which say a lot about the strengths and limits of European digital sovereignty.
The institutional model of DG Translation
On 16 July 2026, the Commission's Directorate-General for Translation (DGT) published the EU Institutional LLM as an open model. It was not designed from scratch: DGT engineers started from existing models by the French company Mistral AI, Mixtral 8x7B and 8x22B, then continued their training with data from Euramis, the vast database of texts translated by professionals for the European institutions. The training relied on EuroHPC's European supercomputers: MeluXina in Luxembourg, Leonardo in Bologna, then MareNostrum 5 in Barcelona for testing.
The problem targeted is very real. In Common Crawl, one of the most widely used corpora for training models, Latvian represents only 0,09 % of the data, Irish 0,07 % and Maltese 0,03 %, according to the Commission. The least represented half of the Union's official languages weighs barely 2,4 % of the total. The DGT therefore guarantees at least one billion tokens to each under-resourced language. On a test bench composed of institutional texts, the announced gains are clear: the Irish score is almost quadrupled, Greek almost doubled, Estonian improves by nearly 80 %, Latvian and Lithuanian by about 70 to 75 %.
The model already powers eSummary, the Commission's multilingual summarization tool. It is distributed under the Apache 2.0 license, but its download, via the European Language Data Space, is reserved for legal entities established in the Union. The DGT has also published EU MMLU, an evaluation benchmark translated and revised by students from the European Master's in Translation network rather than by machine translation.
EUROPA, the ambition of a frontier model
Second project, on a completely different scale: the Frontier AI Grand Challenge, launched in February 2026 with the joint undertaking EuroHPC, invited European players to propose a model with more than 400 billion parameters. On 19 June 2026, the Commission designated the EUROPA consortium, led by the Italian company Domyn, as the sole winner. The future model will be open source and cover the 24 official languages.
"Europe can be at the forefront of advanced AI on its own terms. EUROPA will build a European frontier AI model in the 24 EU languages, showing that we can match the best while remaining true to our values." (Henna Virkkunen, Executive Vice-President of the Commission for Tech Sovereignty, Security and Democracy, 19 June 2026, translated from English)
This project is not an internal Commission tool: it is an industrial bet, which consists of putting public computing power at the service of a European private player so that it can compete with American and Chinese laboratories.
OpenEuroLLM, the path of open research
Third project, OpenEuroLLM brings together twenty European research institutions, companies and computing centers, coordinated by Jan Hajič (Charles University, Prague) and co-directed by AMD Silo AI. Its total budget amounts to 37,4 million euros, of which 20,6 million are provided by the Digital Europe Programme. It aims to create a family of fully open models, including data and training methods, covering all official languages of the Union and other important languages.
After a year of work, the project has published datasets and baseline models, and obtained strategic access to four EuroHPC supercomputers. Its coordinator acknowledges, however, that "significant challenges remain, particularly in obtaining more computing power to create the final models." The consortium aimed for a first model of 8 billion parameters by summer 2026.
Three logics, one same constraint
These three initiatives do not directly compete with each other, but they reveal a still fragmented strategy.
- The institutional model responds to a specific public service need, the multilingual quality of the Commission's tools. It is remarkable that it relies on technology from a private French company: sovereignty here comes from choosing a European provider rather than building from scratch.
- EUROPA bets on scale and competition with global leaders. Its success will depend on effective access to computing capacity and the ability to keep pace with an innovation rate dictated by players with far greater resources.
- OpenEuroLLM prioritizes transparency and reproducibility, core values of the European project, but directly faces the computing shortage.
Computing is the common thread. All three projects depend on EuroHPC supercomputers, which also equip the European 'AI factories'. Yet the power needed to train a frontier model amounts to thousands of graphics processors over months. By comparison, Mistral has indicated it trained its Mistral Large 4 model, presented in October 2026, on about 3 800 Nvidia processors in its own European data centers.
What European law imposes on these models
Contrary to a common misconception, European institutions are not exempt from the AI Act. The regulation even provides for a specific regime of administrative fines for Union institutions, bodies and agencies (Article 100), under the supervision of the European Data Protection Supervisor. For providers of general-purpose AI models, documentation and copyright compliance obligations have applied since 2 August 2025, with derogations provided for models released under a free license, unless they present a systemic risk.
There remains a political question: does the Union primarily want reliable tools for its administrations, a champion capable of competing at the top, or a fully open digital common good? All three objectives are legitimate. But without a clear prioritization of computing power, which is scarce and costly, Europe risks scattering its efforts at a time when its competitors are concentrating theirs.
Sources : European Commission (DGT), publication of the EU Institutional LLM, 16 July 2026 ; DGT, presentation of the EU Institutional LLM ; Slator, 28 July 2026 ; European Commission, selection of the EUROPA consortium, 19 June 2026 ; European Commission, launch of the Frontier AI Grand Challenge ; OpenEuroLLM, first-year review, 6 March 2026 ; European Commission, STEP label awarded to OpenEuroLLM ; Mistral AI, Mistral Large 4 ; AI Act, Article 100.