Europe has reached a new milestone in its quest for technological independence with the launch of EuroLLM-22B, the largest and most advanced open large language model (LLM) developed entirely in Europe.
The model is designed to support all 24 official European Union languages, reinforcing Europe’s commitment to multilingual, open, and sovereign artificial intelligence.
The release was announced today by Técnico, Instituto de Telecomunicações, and the ELLIS Unit Lisbon, as part of a broad European research collaboration.
EuroLLM-22B is the most powerful model in the EuroLLM family, following earlier releases at 1.7B and 9B parameters. Trained from scratch on the MareNostrum 5 at the Barcelona Supercomputing Center, the model delivers performance that is competitive, and in some cases superior, to global models of similar size on multilingual tasks.
By focusing on Europe’s linguistic diversity rather than adapting English-centric systems, EuroLLM-22B sets a new standard for multilingual language models built specifically for European realities.
The model was developed by a consortium of leading academic and industrial partners, including the University of Edinburgh, Paris-Saclay University, University of Amsterdam, Unbabel, Naver Labs, and others. The initiative is supported by Horizon Europe and a EuroHPC Extreme Scale Access grant.
Crucial to the project was Europe’s public supercomputing infrastructure, coordinated by the EuroHPC Joint Undertaking, which currently operates several of the world’s top-ranked supercomputers and is preparing to launch new exascale systems.
A defining feature of EuroLLM-22B is its fully open nature. The model is freely available to researchers, startups, developers, and public institutions, lowering barriers to entry and enabling experimentation across Europe’s innovation ecosystem.
According to André Martins, our senior researcher, ELLIS Fellow and Professor at Técnico, "the project demonstrates what Europe can achieve through open collaboration and shared infrastructure, strengthening long-term AI sovereignty."
As AI development becomes increasingly dominated by major non-European players, EuroLLM positions itself as a strategic alternative. Alongside initiatives such as OpenEuroLLM and Apertus, the project aims to ensure that Europe retains control over critical AI technologies without sacrificing openness or scientific excellence.
Beyond its current multilingual focus, which includes all EU languages and 11 additional ones, the EuroLLM roadmap includes future multimodal capabilities such as speech, vision, and video. These extensions will be powered by training on the upcoming Jupiter exascale system starting in 2026.
With major players like OpenAI, Google, Alibaba, and Meta dominating the AI landscape, reliance on their models poses significant risks, including limited openness and uncertain availability. Together with other European initiatives, such as OpenEuroLLM and Apertus, EuroLLM aims to counter this trend by offering a fully accessible alternative designed to serve Europe’s needs without compromising its independence. As Pierre Colombo, Assistant Professor at CentraleSupelec (University Paris-Saclay), says:
"EuroLLM-22B is the generative counterpart to what we've achieved with encoder models like EuroBERT. Europe now has both precision tools for classification and retrieval, and powerful generative capabilities - all built on our linguistic reality, not as an adaptation of English-centric models. This is what sovereignty looks like: open infrastructure that works for Europe first, not last."
With EuroLLM-22B, Europe signals a clear ambition: to lead in AI development on its own terms, grounded in openness, diversity, and public infrastructure. By uniting world-class researchers and leveraging a continent-wide supercomputing network, the initiative marks a decisive step toward a more independent and inclusive AI future.
*
Languages supported: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Irish, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Arabic, Catalan, Chinese, Galician, Hindi, Japanese, Korean, Norwegian, Russian, Turkish, and Ukrainian.
Developed by: Instituto Superior Técnico - University of Lisbon, Instituto de Telecomunicações, University of Edinburgh, Aveni, Unbabel, University of Paris-Saclay, University of Amsterdam, Naver Labs, Sorbonne Université.
Authors: Miguel Moura Ramos, Duarte M. Alves, Hippolyte Gisserot-Boukhlef, João Alves, Pedro Henrique Martins, Patrick Fernandes, José Pombal, Nuno M. Guerreiro, Ricardo Rei, Nicolas Boizard, Amin Farajian, Mateusz Klimaszewski, José G. C. de Souza, Barry Haddow, François Yvon, Pierre Colombo, Alexandra Birch, André F. T. Martins.
For more information or interview requests, contact: eurollm.team@gmail.com
EuroLLM-22B is available today via Hugging Face, where technical documentation and benchmark comparisons are publicly accessible: