European AI Foundation Model for Industry Unveiled

The Soofi Consortium has unveiled “Soofi S,” the first component of a European AI model family. The open-source base model is aimed at industry, government, and research, and is designed to enable transparent operation on proprietary or sovereign infrastructure.

AI Base Model for Industrial Applications

Soofi S is designed as a technical foundation for industrial AI applications. Unlike a general-purpose chatbot model, it is intended to enable companies to develop their own solutions based on internal data and adapt them to specific processes, document repositories, or software environments.

Intended areas of application include industrial processes, the analysis of extensive technical and regulatory documents, the generation of program code, and agent-based AI systems. The model will initially be released as a base model and can be fine-tuned for specific industries, fields of expertise, or company-specific tasks. Retrained variants for dialogue and agent applications are to follow.

30 Billion Parameters and Hybrid Architecture

According to the consortium, the model was trained from scratch using 27 trillion tokens. It is based on a mixture-of-experts architecture with 30 billion parameters and 3 billion simultaneously activated parameters, referred to as 30B-A3B.

Technically, Soofi S combines Transformer components—which analyze relationships between text segments—with Mamba components, which efficiently process long sequences using a compact internal state. This hybrid architecture is designed to enable high data throughput with comparatively low energy consumption. Training focused on English and German texts.

According to the project consortium, initial benchmark results show that Soofi S can compete with open models of comparable size on English-language tasks. In the German-language benchmarks examined, the model is said to achieve leading results within the comparison group. However, a robust evaluation for specific industrial applications will depend in particular on the respective domain data, fine-tuning, and the infrastructure used.

Training on European Cloud Infrastructure

Soofi S and the planned successor models are being trained on Deutsche Telekom’s Industrial AI Cloud in Munich. NVIDIA Blackwell GPUs and open NVIDIA Nemotron models are used for this purpose.

The project is funded by the Federal Ministry for Economic Affairs and Energy as part of the European IPCEI-CIS and 8ra initiatives and is financed through NextGenerationEU. The consortium is coordinated by the German AI Association.

Participants include Fraunhofer IAIS, Fraunhofer IIS, the German Research Center for Artificial Intelligence, Julius Maximilian University of Würzburg, Leibniz University Hannover, Technical University of Darmstadt, Berlin University of Applied Sciences, as well as Ellamind and Merantix Momentum.

Model weights and training documentation planned

One focus of the project is on technical traceability. In addition to model weights, the consortium plans to provide information on training methodology, data preparation, and the data pipelines used.

This is particularly relevant for companies and public institutions when AI systems need to be audited, adapted to regulatory requirements, or operated on their own infrastructure. System integrators and solution providers thus gain a potential foundation for developing industry-specific AI applications without being entirely dependent on proprietary models from individual non-European providers.

Field Tests with Industry Partners

The model is to be tested in real-world application scenarios in collaboration with industrial companies. The goal is to gather experience from production-like environments and to align further development with specific requirements from the business sector.

For users, a key factor will be how reliably the model can be integrated into existing data architectures, software platforms, and operational processes. Equally relevant are the resource requirements for operation, the quality of domain-specific adaptations, and the long-term availability of technical documentation.

Further technical information, project updates, and contact details for pilot projects are available at https://www.soofi.info/.