Soofi Announces Model for Industrial AI in Europe

Soofi S marks the beginning of the project's first release phase. The model is designed for organizations that need to run AI applications transparently, adaptively, and on their own or sovereign infrastructure — for example, in industrial processes, the analysis of extensive technical and regulatory documents, code generation, or agentic AI systems. Trained from scratch on 27 trillion (27T) tokens, Soofi S is a 30 billion parameter (30B-A3B) mixture-of-experts model whose hybrid Mamba-Transformer architecture combines high throughput with low energy consumption. Soofi S has been trained primarily on English and German text and achieves top-tier results among open models in its size class in English; in German, it leads the peer group. Soofi S is initially released as a base model, which can already be fine-tuned for specific domains; post-trained variants for dialogue and agentic applications will follow.

"Whoever controls the foundation models controls a central part of future digital value creation and strengthens their sovereignty and resilience. With Soofi, we are building an open foundation on which businesses, 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 Center for Sovereign AI, German AI Association.

A particular focus lies on transparency: the consortium will not only release model weights but also publish technical documentation on training methodology, data preparation, and the data pipelines used. This makes Soofi S more auditable for businesses, public authorities, and researchers, and more readily adaptable to specific use cases.

Soofi S and subsequent models are trained on Deutsche Telekom's Industrial AI Cloud in Munich, using NVIDIA's open-source AI framework. Initial results show that the Soofi S base model matches or outperforms international models of comparable size across German and English benchmarks.

"Soofi S is not intended as yet another general-purpose chatbot, but as a technical foundation for industrial AI. What matters is that Soofi S performs not only well in benchmarks, but can be deployed reliably, efficiently, and transparently in production," says Nicolas Flores-Herr, Technical Project Lead for Soofi and Team Lead at Fraunhofer IAIS.

The model is being tested jointly with industrial partners in real-world application scenarios. The aim is to gather early experience from practical deployment and to align the model's further development closely with concrete industry requirements, enabling companies to integrate the model into their processes and products. 

Organizations interested in piloting or collaboration are welcome to get in touch at contact@soofi.info.