German research group unveils 'Soofi S', a 30B open-source LLM outperforming larger models in English and German
A consortium of German research institutes and companies has launched Soofi S, an open-source bilingual AI model that uses novel techniques to achieve top-tier benchmark scores.
A German AI consortium, coordinated by the German AI Association (KI Bundesverband), has unveiled "Soofi S 30B-A3B," an open-source Large Language Model (LLM) that has stunned the community by topping benchmarks for fully open models in both English and German, surpassing leaders like OLMo 3 32B and Apertus 70B.
Soofi S’s standout feature is its hybrid architecture combining Mamba and Transformer, utilizing a Mixture-of-Experts (MoE) technique. Despite a total parameter count of 31.6 billion, the MoE approach ensures only about 3.2 billion parameters are active per token, enabling high speed and efficiency without sacrificing large-scale capabilities.
This project is a collaboration between institutions such as Fraunhofer Institutes, DFKI, the Technical University of Darmstadt, and the AI firm Ellamind, funded by a €20 million grant from the German Federal Ministry for Economic Affairs and Climate Action under the European IPCEI-CIS initiative. The model was trained from March to May using up to 512 Nvidia B200 GPUs at the Deutsche Telekom data center in Munich, totaling over 253,000 GPU-hours.
In coding tasks, Soofi S achieved an impressive 73.8% on HumanEval and 70.2% on MBPP—strong figures for a general-purpose language model. This launch marks a significant milestone in demonstrating Europe's potential to develop native foundational AI technology to reduce reliance on foreign tech.
The emergence of Soofi S proves that national collaborations can produce high-quality AI models that rival big tech companies, providing a vital open-source alternative for developers worldwide, including those in Thailand.