Artificial Intelligence & European Technology Policy

SOOFI: Europe's Bid to Build Its Own Open Source AI

A €20 million project uniting six European research centers will train a 100-billion-parameter, open-source language model designed to be trusted — and governed — in Europe.

• About €20 million from Germany's Federal Ministry for Economic Affairs and Energy funds the consortium.
• Six research institutions plus two start-ups are involved: L3S, Fraunhofer IAIS, Fraunhofer IIS, DFKI, University of Würzburg, TU Darmstadt, Beuth University, coordinated by the German AI Association.
• The goal is a model of roughly 100 billion parameters, open source, multilingual, and aligned with European ethical and regulatory values.

The project, SOOFI — Sovereign Open Source Foundation Models — is built on a simple observation: Europe already owns a lot of data, but almost none of the foundation models that run on it. Today, the large language models underpinning research, industry, and public-sector software are supplied almost entirely by non-European companies. It is the same dependency pattern that the continent learned from the cloud era.

SOOFI's answer is to train the model in-house and publish it openly. The target is a 100-billion-parameter base language model, with a specialized reasoning model stacked on top. That second piece matters: reasoning models can chain steps, consult outside sources, and support agent systems that automate real workflows in manufacturing, law, and healthcare.

L3S, one of the lead partners, handles multilinguality, safety, and value alignment — developing fine-tuning datasets, safety benchmarks, and reward models so the system behaves reliably across languages and respects cultural context. Director Wolfgang Nejdl puts it plainly: models in education and medicine need not just technical excellence but cultural and ethical alignment.

The honest caveat is scale. A hundred billion parameters sits below the very largest frontiers of 2026, so SOOFI is not chasing raw leaderboard supremacy. It is building an auditable, open, regionally trusted alternative — the kind of infrastructure that matters most when the model itself is the foundation for everything else.