German Consortium Releases Soofi S, Open 30B Model Topping Benchmarks
Key Takeaways
- ▸Soofi S is the first open-source model trained entirely on Deutsche Telekom's Industrial AI Cloud, achieving top benchmark performance across English, German, and programming tasks
- ▸The MoE architecture activates only 3.2B of 31.6B parameters per token, achieving 8x higher throughput than dense models while maintaining constant performance at context lengths up to 256,000 tokens
- ▸The model challenges classical Chinchilla scaling laws by training on 900+ tokens per parameter—a ratio defended by project leads as appropriate for MoE architectures where data repetition benefits individual experts
Summary
The KI Bundesverband (German AI Association) has released Soofi S, an open-source 30-billion parameter language model that achieves the highest scores on English and German benchmarks among fully open models, surpassing competitors like OLMo 3 32B and Apertus 70B. The model was trained entirely on Deutsche Telekom's Industrial AI Cloud infrastructure using 27 trillion tokens, with training data deliberately weighted toward German-language content to address the multilingual gap in open-source LLMs.
Soofi S employs a mixture-of-experts (MoE) architecture that activates only 3.2 billion of its 31.6 billion total parameters per token, delivering computational efficiency comparable to traditional 3B models. The hybrid design combines Mamba-2 layers with standard attention mechanisms, with only 6 of 52 layers maintaining a key-value cache. This architecture delivers approximately eight times the throughput of dense 14-24B models on long contexts, maintaining nearly flat performance from 4,000 to 256,000 tokens—a significant advantage for processing lengthy documents.
The launch has sparked debate within the AI community about training efficiency and scaling laws. Critics argue that Soofi S violates classical Chinchilla scaling principles by using ~900 tokens per parameter, far exceeding the 20 tokens-per-parameter ratio recommended by DeepMind in 2022. Project technical lead Michael Fromm countered that these classical scaling laws apply only to dense models, not MoE architectures, where individual experts benefit from seeing repeated high-quality data—a principle Nvidia itself has adopted with models trained on up to 25 trillion tokens.
- As a German-focused, fully open model, Soofi S strengthens European AI autonomy and addresses the multilingual gap in open-source LLMs dominated by English-centric alternatives
Editorial Opinion
Soofi S represents a meaningful step forward for European AI independence and open-source model development outside the U.S. tech sphere. The MoE architecture's efficiency gains are impressive and suggest that classical scaling laws may require significant revision for modern architectures—a debate that extends far beyond this single model release. However, the overtraining controversy also highlights the urgent need for more rigorous, architecture-aware evaluation frameworks; the community should invest in better benchmarking standards to move beyond scaling-law debates and toward truly predictive design principles.
