AI & Automation

Sovereign AI from Europe: Kolibri, Soofi, Mistral and more at a glance

Which open AI models from Germany and Europe exist in 2026, what sets them apart and how to run them without US providers.

22 October 2026 · Levisteria

2026 is the year in which serious open AI models from Germany arrived. With Kolibri from Aleph Alpha and Soofi S from a research consortium, two German-English language models with open weights have been released. Add to that models from France, Switzerland and EU projects. Anyone wanting to run AI without depending on US providers now has a real choice. The top capabilities still belong to the large models from the US and China, but for many company tasks the European models are enough.

Which models are there?

  • Kolibri (Aleph Alpha, Heidelberg): since October 2026, 78 billion parameters, around 3.5 billion of them active per word, weights under Apache 2.0. Strong in German, built for data-center GPUs. More in our article Kolibri.
  • Soofi S (consortium, trained in Deutsche Telekom's Industrial AI Cloud in Munich): since July 2026, around 31 billion parameters, about 3 billion of them active, with a stronger weighting of German. Particularly open: besides the weights, details of the training data as well as training and evaluation code are to be published.
  • Mistral (France): the best-known European provider, several models with open weights, including large models under Apache 2.0.
  • Apertus (Switzerland): from ETH Zurich, EPFL and the Swiss National Supercomputing Centre, with disclosed architecture, weights, training data and method.
  • EuroLLM: an EU-funded model covering all official EU languages.
  • Teuken (OpenGPT-X): an earlier German project that Soofi is meant to replace.
  • FLUX (Black Forest Labs, Freiburg): not a language model but image AI, some with open weights, more in AI images in 2026.

Versions and licences change quickly. Check the provider's current model card before use.

What does "sovereign" actually mean?

The term is used in different ways. It can mean the model was developed in Europe, that its weights are freely available, that training data is disclosed or that it runs on your own infrastructure. For your company, the last point matters most: if the model runs on a server you control, your data never leaves the building, wherever the model comes from.

Where can you run the models?

  • On your own servers: full control, but you need suitable hardware, more in Local AI in practice.
  • With European cloud providers: IONOS, OVHcloud or STACKIT run open models in European data centers.
  • Mixed: sensitive tasks locally, non-critical ones with an EU provider.

What happens next?

Politics is backing the topic: on 19 October 2026, Germany's economics and research ministries signed a letter of intent with Nvidia for open, specialised AI models for industry. For mid-sized companies, this means the choice of European models will grow and running them in Germany will get easier.

Sources

← All articles

Let’s connect

AI without depending on US providers?

Together we choose a suitable European or open model and run it on your server or in a German data center.

30 min · free · an honest take on whether it's worth it

What happens3 steps
first call — 30 minfree
  1. You tell us where things get stuck.
  2. We ask questions and look at your workflows.
  3. You get an honest assessment and, if it fits, a suggestion for the next step.
→ Then: a concept with clear effort and benefit
Open source tools we use

Accessibility

Adjust the page. Your choice only applies on this device.

Accessibility statement →