With SOOFI (Sovereign Open Source Foundation Models), one of the most ambitious initiatives to strengthen European AI sovereignty is now underway. The project aims to develop an open AI language model with around 100 billion parameters that meets European values and regulatory requirements. Six leading research institutions are involved – including the L3S Research Center.
Why SOOFI?
Europe currently has very few large-scale language models of its own that can serve as foundations for sector-specific applications. Powerful reasoning models capable of solving complex tasks through structured thinking are equally scarce. As a result, dependence on non-European providers continues to grow – a pattern already familiar from the cloud sector. SOOFI aims to create a viable alternative.
Building a European AI language model will reduce these dependencies and provide a reliable basis for innovative and trustworthy applications in industry and the public sector. In addition, a specialised reasoning model will be developed on top of the language model to improve system quality and handle more complex tasks. Initial application scenarios using AI agent technologies are also planned.
Reasoning models are of particular importance for German industry: they analyse complex technical, regulatory and organisational contexts, draw on additional information sources when needed, and support well-founded decision-making. They enable new forms of automation using interconnected agent systems and contribute to quality improvements in development, production and knowledge management.
L3S: Multilinguality, Safety and European Values
Within the project, L3S is responsible for key tasks related to multilinguality, safety and value alignment. Researchers are developing multilingual datasets for fine-tuning the models, creating safety benchmarks and working on reward models and reasoning datasets to strengthen logical capabilities across different languages.
L3S also ensures that the models reflect European cultural and ethical values, and is building a comprehensive evaluation platform that systematically compares SOOFI models with existing systems. Additional priorities include identifying relevant use cases and adapting the technology for concrete applications.
“Large language models that respect European values are essential for building trust in AI — especially in sensitive domains such as education and medicine. These areas require not only technical excellence but also cultural and ethical alignment to ensure that AI supports people responsibly and transparently,” says L3S Director Professor Wolfgang Nejdl.
SOOFI thus represents not only a technological foundation for European industry, but also a clear commitment to open, value-driven AI systems.
Consortium and Funding
SOOFI is funded with approximately €20 million by the Federal Ministry for Economic Affairs and Energy. Project partners include L3S, Fraunhofer IAIS, Fraunhofer IIS, DFKI, the University of Würzburg, TU Darmstadt, the Beuth University of Applied Sciences Berlin, and two start-ups. The consortium is coordinated by the German AI Association.
