German AI Association Releases Open-Source Model Soofi S, Leading in English and German Benchmarks
German AI consortium releases Soofi S, an open 30B model that tops benchmarks in both English and German
AI BriefingA research consortium coordinated by the German AI Association released the open-source large language model Soofi S 30B-A3B. The model has 31.6 billion total parameters, with only approximately 3.2 billion parameters activated per token, and uses a MoE architecture that combines Mamba-2 with standard attention layers. The model was trained entirely on Deutsche Telekom’s Munich Industrial AI Cloud in Germany. The share of German in the training data increased from 7.2% in the first phase to 15.3% in the second phase. In benchmark tests, Soofi S achieved the highest combined English and German scores among all fully open-source models, surpassing OLMo 3 32B and Apertus 70B. It scored 73.8% on HumanEval, 70.2 on MBPP, and 84.2 on the German version of MBPP. The context window supports up to 1 million tokens. At a length of 40,000 tokens, its generation throughput is approximately 8 times that of dense models of comparable size. The model weights have been open-sourced.
Also reported by 1 other source
Source: The Decoder: AI News · the-decoder.com
Event follow-up1
View all event coverage- German AI Association Releases Open-Source Model Soofi S, Leading in English and German BenchmarksThe Decoder: AI News · July 13