SOOFI. A Sovereign, Open-Source Foundation Model.

Soofi-S is a compact 30-billion-parameter model specifically designed for locally running AI agents. It is intended to run on a powerful PC or Mac with a single consumer GPU, handling tasks such as tool and function calling, multi-step workflows, coding, file processing, and personal assistance. Its primary purpose is to enable powerful AI agents to run directly on users’ own devices, including offline and with sensitive personal data that does not necessarily need to be transferred to the cloud.

The role of the BHT team at DATEXIS is mainly on enabling complex reasoning. Reasoning is essential because science and other complex business tasks rarely works with perfect, standardized datasets. In their daily routine, users must combine incomplete evidence from different sources, judge the strengths and weaknesses of different methods, and continuously revise their hypotheses as new evidence emerges. AI agents can mimic this process by using reasoning to choose tools, evaluate results, criticize hypotheses, and decide what to investigate next. In this sense, reasoning enables AI to reproduce the iterative and uncertain nature of real scientific discovery.

Soofi S: The first model generation of the Soofi project. The model was developed and trained in Germany based on a specially curated data corpus and forms the technological basis for further model generations.

The focus is on a balanced relationship between quality, computational effort, speed, and operating costs. As a foundation model, Soofi S is designed for diverse application areas, including industrial applications, knowledge management, software development, regulatory tasks, and agentic AI systems.

Soofi S is currently in the beta testing phase with selected partners and users. Further applications for the testing phase cannot be accepted at this time.

The public open-source release is planned for this year.