The bottom line: Soofi S activates only 3 billion out of 30 billion parameters per token, surpassing European and US open-source models with significantly lower computational requirements.
The Soofi research consortium has released the open-source model Soofi S 30B-A3B, which surpasses international open-source competitors in English and German benchmarks. The model combines Mamba and Transformer architectures with a Mixture-of-Experts design and was trained on Deutsche Telekom’s AI cloud infrastructure.
The Soofi S 30B-A3B model is based on a Mixture-of-Experts (MoE) architecture, in which the neural network is divided into specialized sub-units that are selectively activated for specific data packets. The overall architecture combines Mamba and Transformer elements. With 30 billion total parameters, the system activates only 3 billion parameters per processed token. This significantly reduces computational resource requirements during operation.
An advantage of the MoE architecture lies in memory efficiency for context: the cache for computations remains nearly constant as text context grows. This enables speed advantages when processing long documents and handling simultaneous requests from multiple users compared to densely packed Transformers. The application was trained entirely on Deutsche Telekom’s AI cloud infrastructure.
According to a published technical report on arXiv, Soofi S is currently the most powerful fully open model in English and German test procedures. It surpasses open-source competitors such as Alia 40B (Spain), EuroLLM 22B, Apertus 70B (Switzerland), and Olmo 3 32B (USA). In aggregated tests, Soofi S achieves the performance of closed models with 14 to 27 billion parameters, but requires only a fraction of the active parameter costs. The model performed weaker on MMLU-Pro (general knowledge) and GPQA-D-DE (advanced scientific reasoning), but also outperformed other open-source comparison models there. Michael Fromm, Head of Pretraining Data at Soofi, emphasized: “Our advantage is performance + throughput + complete openness.”
Access to the model is not implemented as a freely available chatbot on the internet. Companies and organizations must submit access requests with a specific use case. Karl Lauterbach, Chair of the Committee on Research, Technology and Space in the Bundestag, stated that Soofi does not compete with top-tier models like Claude or OpenAI systems, but is sufficient for a broad spectrum of industrial applications and administrative tasks.
Source: www.it-daily.net · Published 17 July 2026
Lumi AI News — AI-assisted curation pursuant to Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.