Soofi S marks the start of the project’s first release phase. The model is intended for use in particular where organisations wish to operate AI applications in a traceable and customisable manner on their own or sovereign infrastructure – for example, in industrial processes, the analysis of extensive technical and regulatory documents, code generation or agent-based AI systems. Trained from scratch on 27 trillion (27T) tokens, Soofi S – a 30-billion-parameter (30B A3B) mixture-of-experts model – combines high throughput with low energy consumption thanks to its hybrid Mamba Transformer architecture. Soofi S has been trained primarily on English and German texts and, in English, achieves top scores among open-source models in its size class; in German, it leads the comparison group. Soofi S is initially being released as a base model that can already be fine-tuned for specific domains; retrained variants for dialogue and agent applications will follow.

“Whoever controls the base models controls a central part of future digital value creation and increases their autonomy and resilience. With Soofi, we are creating an open foundation on which companies, SMEs and the public sector can develop transparent AI applications based on their own data, without becoming permanently dependent on individual non-European models,” says Jörg Bienert, Managing Director of the Centre for Sovereign AI at the German AI Association.

There is a particular focus on transparency: the consortium will not only provide model weights but will also publish technical documentation on training methodology, data preparation and the data pipelines used. This will make Soofi S more verifiable for companies, public authorities and researchers, as well as adaptable to specific fields of application. TU Professor Kristian Kersting, co-spokesperson for the Cluster of Excellence Reasonable Artificial Intelligence (RAI), and his team (Ruben Härle, Lukas Helff, Maurice Kraus and Sebastian Sztwiertnia) are involved in this project. Specifically, they are engaged in three tasks: the researchers will set up an innovative data pipeline designed to collect reliable, European training data with the aid of AI-supported quality checks (“LLM Judges”). In addition, the researchers are developing and training a reasoning model capable of drawing better logical conclusions. Such systems could, for example, make administrative processes more comprehensible, support craft businesses in preparing quotations, or assist start-ups with technical decisions. Thirdly, the group is working on energy-efficient alternatives to traditional transformers to enable more cost-effective AI services in the long term.

Soofi S and subsequent models are being trained on Telekom’s Industrial AI Cloud in Munich, using NVIDIA Blackwell GPUs and NVIDIA Nemotron open AI models. Initial results show that the Soofi S base model can keep pace with – or even outperform – international models of comparable size in German and English benchmarks. SOOFI/mho