EuroLLMUnclaimed

EuroLLM is a family of large language models built in Europe to cover all 24 official EU languages. The models are trained on multilingual data across 35 languages, and the team says they are designed to support the EU's goal of AI sovereignty.

The current flagship, EuroLLM-22B, is a 22 billion parameter model trained on over 4 trillion tokens and supports a context window of 32K tokens. Two versions are offered: an Instruct model fine-tuned for instruction following and chat, and a Base model intended for fine-tuning on any task. A 9B parameter model is available in Base and Instruct versions, and a 1.7B model is described as suited to edge devices.

The page also lists three models that are coming soon: EuroVLM-9B, a multimodal vision-language model that adds a vision encoder to EuroLLM-9B, a sparse mixture-of-experts model with 600M active parameters called EuroMoE-2.6B-A0.6B, and vision and voice support for the wider model family.

The models are aimed at researchers, organisations and citizens in Europe who want an open model to use or build on. The page says the models are freely used and fully open source, and they are available through Hugging Face, where the team links to model pages and release articles. The site describes the models as strong on question answering, summarisation and translation.

The project is run by a consortium of universities and research groups, including Instituto Superior Tecnico, the University of Edinburgh, Universite Paris-Saclay, Unbabel and the University of Amsterdam. The page says it was supported by Horizon Europe, the European Research Council and EuroHPC, and that it was trained on the MareNostrum 5 supercomputer.

The page does not list prices. The models are presented as free and open source.