German consortium releases Soofi S, the top-ranked open-source AI model
Germany enters the AI race.
Image by Cybernews
- German researchers released Soofi S, an open-source German-English AI model that led all tested open-source benchmarks.
- The model uses 30 billion parameters but activates only 3 billion per token, reducing computing needs.
- Soofi says the model outperformed European and US open-source rivals in English, German, coding, maths, and regional knowledge.
- German observers say Soofi S could support more independent AI use in industry and public administration.
Key Takeaways by nexos.ai, reviewed by Cybernews staff.
A German research consortium, Soofi, has released an open-source large language model (LLM) called Soofi S, which outperforms its open-source counterparts across all tested benchmarks.
Soofi S 30B-A3B is a sovereign, open-source Mixture-of-Experts (MoE) hybrid Mamba-Transformer foundation model for German and English.
MoE refers to a machine learning approach that divides an AI model into separate sub-networks, also known as “experts.” Each of them specializes in a subset of the input data, enabling them to jointly perform a task.
The architecture is used in all 10 of the most intelligent open-source models, including DeepSeek AI’s DeepSeek-R1 and OpenAI’s gpt-oss-120B, according to Artificial Analysis, an independent AI benchmarking platform.
Soofi S contains 30 billion parameters, but activates only 3 billion per token, which requires less compute.
In addition, Soofi S keeps the inference cache – storing of previous results – near-constant as context grows, giving it “throughput advantage over dense models for long-context, high-concurrency deployment.”
Dense models, such as Claude Opus 4.5 and Gemini 3.5 Flash, are those where every token passes through the same feed-forward network in every transformer block.
Soofi S was fully trained on Deutsche Telekom's AI cloud infrastructure, a German telecommunications company.
The best open-source AI model?
According to the pretraining tech report published on the preprint server arXiv, Soofi S is the strongest fully open model on English and German benchmarks, outperforming Spain’s Alia 40B, the joint European EuroLLM 22B, Switzerland's Apertus 70B, and the US’s Olmo 3 32B.
It also demonstrates strong performance in coding, maths, and regional knowledge of German.
The paper suggests that Soofi S matches dense 14-27B international models on English and German aggregate performance at “a fraction of their active-parameter cost.”
Michael Fromm, head of pretraining data at Soofi, later added that Qwen 3.5 is ahead of Soofi S in terms of raw capability.
“Our edge is capability + throughput + full openness,” he wrote on X.
Soofi S demonstrated the weakest performance on the model’s acquired knowledge and reasoning skills (MMLU-Pro) and highly advanced scientific reasoning (GPQA-D-DE) benchmarks.
However, it still outperformed other open-source models included in the analysis.
“Germany is back in the AI race”
Daniel D. Eckert, a reporter at the leading German daily Welt, wrote on X that Soofi S’s release may mean that Berlin is back in the AI race.
“We would have our own independent AI technology and would no longer be dependent on American or Chinese companies,” Eckert wrote.
Karl Lauterbach, a chairman of the Committee on Research, Technology, and Space at the German parliament, said that Soofi doesn’t compete with Claude or OpenAI's top models.
“But for a very wide range of industrial applications and administration, it's more than good enough,” Lauterbach wrote on X.
How to access the Soofi S model?
Soofi S isn’t available as a chatbot, but enterprises can submit a request for access, which requires them to provide an application scenario.
