Soofi: Sovereign AI from Europe​

 Soofi (Sovereign Open Source Foundation Models) is a German consortium project for the development of sovereign , open AI foundation models for Europe. Its goal is to offer companies , SMEs , public administration, research institutions, and startups transparent and independent alternatives to non-European AI models .  
 

Soofi is supported by a consortium of leading German research institutions , universities , and AI startups : the German Research Center for Artificial Intelligence (DFKI), the Fraunhofer Institute for Intelligent Analysis and Information Systems (IAIS), the Fraunhofer Institute for Integrated Circuits (IIS), TU Darmstadt, the University of Würzburg, Berlin University of Applied Sciences, the L3S Research Center of Leibniz University Hannover, and the Lamarr Institute for Machine Learning and Artificial Intelligence . The startups ellamind and Merantix Momentum are also involved . The consortium is led by the German AI Association . 

"Soofi S" is the first model in the Soofi model family: a 30-billion-parameter model with a hybrid Mamba Transformer architecture as a Mixture of Experts ( MoE ), in which only around three billion parameters are active per token . It was trained with 27 trillion tokens , primarily in German and English. The base model will be released first.  

A base model has undergone pre-training and serves as a starting point for custom adaptations . This is precisely where its value lies : science, industry, and the public sector can build their own models for their specific domain and tasks upon it . Those who begin with the base model retain full control over every subsequent step. For research and domain adaptation, the base model is the crucial publication. 

An instruction-optimized model (post-training) will follow at a later date . It is aimed at all those who want to use Soofi S. Both variants form the foundation for further , larger Soofi models, including specialized variants for dialogue, reasoning, and agent applications. 

Soofi S utilizes the openly published reference architecture of Nemotron 3 Nano. This allowed us to train a high-performance model in just a few months. In parallel, we are researching our own architectures as a further building block of a sovereign European model stack. The Mixture of Experts (MoE) means that the model has a total of 30 billion parameters , but only activates a small subset of them (around 3 billion) for each query. This significantly reduces computational effort and hardware requirements while maintaining performance. In combination with the hybrid Mamba Transformer architecture, the processing speed remains almost constant even with very long texts, making Soofi S particularly efficient for industrial applications. 

Soofi follows a consistently open approach: In addition to the model weights, selected intermediate checkpoints, the training and evaluation code, and a detailed data inventory are published. The Soofi model family will be released under a permissive license after completion, testing by companies, and passing all required release gates. The conscious decision for an open-source approach is a central element of the project's sovereignty strategy. 

Soofi S was trained on Deutsche Telekom's Industrial AI Cloud in Munich, which was built in partnership with NVIDIA and launched in February 2026. We utilized up to 512 NVIDIA B200 GPUs – 64 DGX B200 nodes, each with eight GPUs – for approximately 253,000 GPU hours between March 24 and May 13, 2026. This made Soofi S one of the first large-scale training runs on this new sovereign HPC-scale infrastructure.

The Munich data center is powered entirely by renewable energy, cooled with water from the nearby Eisbach stream, and integrated into a waste heat recovery system that also supplies heat to the surrounding Tucherpark industrial park . 

Soofi S is currently being tested as part of beta testing within the IPCEI-CIS/8ra initiative – a European project with over 100 partners from industry, SMEs, and research institutions in twelve EU member states. The release will follow as soon as possible.  

For Soofi, sovereignty means that European organizations can use AI technology without becoming permanently dependent on individual non-European providers. Specifically, this means that Soofi S was trained entirely on European infrastructure (Deutsche Telekom's Industrial AI Cloud in Munich), the training data and methods are documented and traceable, and the model can be run on the organization's own hardware. This allows companies and public authorities to retain control over their data, their value creation, and their technological independence. 

"Soofi S" utilizes the openly published reference architecture of  Nemotron  3 Nano. This allowed us to train a high-performance model in just a few months. In parallel, we are researching our own architectures as a further building block for a sovereign European model stack. 

The German Federal Ministry for Economic Affairs and Energy is funding Soofi with approximately 20 million euros. This funding is provided within the framework of the European IPCEI-CIS/8ra initiative and is co-financed by funds from the European Union ( NextGeneration EU).