Release of EuroLLM-22B, a large, fully open-source language model, trained in Europe
Today, the European #EuroLLM consortium announces the release of EuroLLM-22B, a large, fully open-source language model trained in Europe and covering all 24 official languages of the European Union.
This innovation was co-designed at MICS laboratory of CentraleSupélec by Hippolyte Gisserot-Boukhlef, CIFRE doctoral student at Artefact Research Centerwith the contribution of Nicolas Boizard, CIFRE PhD student at Diabolocom, under the supervision of Pierre Colombo et Céline HudelotThey worked hand in hand with the team from theHigher Technical Institute from Lisbon, in particular Miguel Moura Ramos and Duarte Alves, key players in the project supervised by André Martins.
With 22 billion parameters, EuroLLM-22B sets a new standard for multilingual models: competitive performance, or even superior performance, to comparable size international industry models, while being designed from the ground up for European linguistic diversity.
- EuroLLM-22B covers the 24 official EU languages + 11 additional languages, and will expand from 2026 to multimodal capabilities (text, speech, vision, video) thanks to exascale access on the Jupiter supercomputer.
- Open-source by design, EuroLLM-22B can be freely used, studied, and adapted by researchers, startups, SMEs, and public institutions. The goal: to reduce dependence on closed, non-European models and create a genuine ripple effect for innovation in Europe.
- Trained from scratch on the MareNostrum 5 supercomputer at the Barcelona Supercomputing Center, supported by Horizon Europe et EuroHPC, a strategic pillar of the European high-performance computing infrastructure, and by a large European academic and industrial consortium including theInstituto Superior Técnico de Lisbon Telecommunications Institute University of Edinburgh, CentraleSupélec – Paris-Saclay University, Sorbonne University University of Amsterdam, Naver Labs, UnbabelAveni, Artefact Research Center et Diabolocom.
As Pierre Colombo (Centrale Supélec - Paris-Saclay University), “Europe now has both comprehension models (like EuroBERT) and powerful generative models, built on our linguistic reality — that is digital sovereignty.”
EuroLLM-22B is available today at Hugging Facewith detailed results on public benchmarks. A key step towards an open, inclusive European AI aligned with our values.
