EU releases multilingual AI model for all 24 official languages
DGT has released an open AI model built to treat all 24 EU languages equally.
The European Commission’s Directorate-General for Translation (DGT) has released an open large language model designed to provide balanced performance across all 24 official EU languages. Named the EU-Institutional-LLM, it is available in both base and instruct versions through the European Language Data Space.
Unlike many large language models that perform best in only a handful of languages, the EU-Institutional-LLM was designed to provide balanced support across all official EU languages while understanding the legislation, terminology and policy context specific to the Union. The model is trained using DGT’s high-quality multilingual datasets.
The base model is a continually pretrained version of Mixtral-8x7B-v0.1 with around 47 billion parameters and a Mixture-of-Experts architecture. It was trained to improve multilingual performance while limiting catastrophic forgetting through replay data. The instruct version was then created through supervised fine-tuning followed by preference alignment using odds ratio preference optimisation.
Training relied on European supercomputing infrastructure, including the Leonardo Booster system through the EuroHPC Joint Undertaking (EuroHPC JU), together with the MeluXina and MareNostrum supercomputers in Luxembourg, Italy and Spain.
Why does it matter?
The release addresses a longstanding challenge in multilingual AI. Most frontier language models are optimised for a relatively small number of widely spoken languages, creating uneven performance across multilingual societies. By designing a model specifically for all 24 official EU languages and embedding knowledge of EU legislation and institutions, the Commission is seeking to make AI more accessible and useful throughout the Union’s public sector.
The project also illustrates Europe’s broader strategy of building digital sovereignty through open AI infrastructure. Rather than relying exclusively on proprietary foreign models, the EU is investing in publicly available models, European datasets and domestic supercomputing resources that can support research, translation, public administration and future AI applications aligned with European values and regulatory frameworks.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!