LISBON, Could 28, 2025 | Multilingual open-source initiatives EuroLLM and OpenEuroLLM have joined forces to safe 3 million GPU hours on Leonardo – one in every of Europe’s strongest supercomputers – to develop a groundbreaking artificial dataset protecting 40 European languages.
The initiative was chosen underneath the EuroHPC AI Manufacturing facility Massive Scale name recognizing its potential to advance Europe’s management in multilingual synthetic intelligence.
On the coronary heart of this initiative is a mission to construct strategic autonomy for Europe in AI growth. By producing high-quality, ethically sourced artificial information, it addresses a long-standing hole in linguistic illustration, specifically for low-resource and minority languages.
André Martins, Chief Scientific Officer at Unbabel and EuroLLM mission co-lead stated:
“By becoming a member of forces by EuroLLM and OpenEuroLLM, we’re bringing collectively the analysis energy and open-source ethos wanted to sort out one in every of Europe’s greatest AI challenges: linguistic inclusion at scale. This mission is about guaranteeing Europe owns its language information, displays its cultural variety, and units its personal requirements in accountable AI growth.”
The GPU allocation will energy the MultiSynt strategy, a key element of the mission which seeks to deal with one of the crucial persistent bottlenecks in multilingual LLM growth: the shortage of high-quality pre-training information.
“This is a vital step in securing giant sufficient computing energy to construct the OpenEuroLLM’s household of open LLMs. I’m additionally glad that this has been finished in collaboration with the skilled crew from the EuroLLM mission. The objective of this subproject is to discover multilingual artificial information creation and consider their use to be able to attain the next widespread objective: constructing high-quality multilingual LLMs for all European languages and past.” – notes Jan Hajic, Charles College, coordinator of the OpenEuroLLM mission.
Whereas most artificial information technology for giant language fashions thus far has targeted on English, MultiSynt will create the primary complete multilingual artificial dataset designed particularly for pre-training. By leveraging generative fashions to boost and diversify current content material, it can help the broader goals of EuroLLM and OpenEuroLLM: constructing open-source, culturally grounded, and linguistically numerous AI for Europe.
This technique will help linguistic variety, open entry, and information high quality and aligns with the broader aims of the European Fee’s Digital Decade and the AI Act.
The awarded 3 million hours replicate a powerful endorsement of the mission’s technical advantage and strategic worth.
The initiative can be executed by phased releases of the artificial dataset.
****ENDS****
About EuroLLM
The EuroLLM mission consists of Unbabel, Instituto Superior Técnico, the College of Edinburgh, Instituto de Telecomunicações, Université Paris-Saclay, Aveni, Sorbonne College, Naver Labs, and the College of Amsterdam. Collectively they created EuroLLM-9B, a multilingual AI mannequin supporting all 24 official EU languages. Developed with help from Horizon Europe, the European Analysis Council, and EuroHPC, this open-source LLM goals to boost Europe’s digital sovereignty and foster AI innovation.
About OpenEuroLLM
Bringing collectively 20 of Europe’s main AI firms, analysis establishments and EuroHPC centres, the OpenEuroLLM mission is creating a brand new technology of open supply giant language fashions for European languages. Co-funded by the European Union’s Digital Europe Programme, the mission is laying the foundations for AI infrastructure that can improve competitiveness, resilience, and digital sovereignty.
About EuroHPC
The European Excessive Efficiency Computing Joint Endeavor (EuroHPC JU) is a joint initiative between the EU, European nations, and personal companions to develop a world-class supercomputing ecosystem in Europe.
Media Contacts:
For extra info or interview requests, please don’t hesitate to achieve out to our media contacts beneath:
• Unbabel: [email protected]
