German AI Consortium Releases Soofi S — An Open 30B Model That Tops Benchmarks in Both English and German
A German AI consortium released Soofi S, a 30-billion parameter open-source model using Mixture-of-Experts with only 3.2B active parameters per token, topping benchmarks in both English and German while championing transparency.
Europe's New AI Milestone
In July 2026, the German AI consortium unveiled Soofi S — a 30-billion parameter open-source language model that tops benchmarks in both English and German. This marks one of the most significant open-weight releases from Europe, involving Fraunhofer IAIS, Ellamind, and other German research institutions. The model represents Europe's push for AI sovereignty through transparent, high-quality open-source development.
Mixture-of-Experts Architecture
Soofi S utilizes a Mixture-of-Experts (MoE) architecture, meaning that despite its 30 billion total parameters, only approximately 3.2 billion parameters are active per token. This design enables strong performance while keeping computational costs manageable. The model was trained on an impressive 27 trillion tokens, significantly exceeding the conventional Chinchilla scaling laws' recommended ratio of roughly 20 tokens per parameter. Soofi S achieves a ratio of several hundred to one — with only the 3.2B active parameters factored in, the ratio jumps to several thousand to one.
Benchmark Leadership
Soofi S demonstrated top-tier results across both English and German benchmarks, outperforming other open models of comparable size. Its unique bilingual training approach makes it particularly effective for German-language tasks while maintaining competitive English performance. The consortium's transparent evaluation methodology — publishing roughly 152,000 individual results for independent verification — sets a new standard for open model accountability.
The GPQA Contamination Incident
The consortium published version 3.0 of its pretraining report, documenting a data contamination incident involving the GPQA benchmark. The problem occurred because GPQA on Hugging Face has no separate training set — all test material sits under the default 'train' label, causing test questions to mix in with practice data during training. The consortium responded transparently: removed GPQA from evaluation, recalculated all results across 16 compared models (ranking order remained unchanged), and implemented automated cross-checks against all test questions to prevent recurrence.
AI Sovereignty Through Open Source
Soofi S represents a significant step toward European AI sovereignty. While American and Chinese models dominate the market, Europe is actively building its own open AI ecosystem. The open-source approach means any business — including those in Georgia — can use Soofi S for their own products, fine-tuning, or research without vendor lock-in. This democratization of AI access is particularly valuable for smaller markets and non-English speaking regions.
What's Next: Soofi L
The consortium is already training Soofi L, a larger variant. Fine-tuned instruct and reasoning variants are currently in beta testing and expected to ship under a permissive license in the coming weeks. The team's commitment to transparency and community-driven development positions Soofi S as a compelling alternative to proprietary models for developers and organizations worldwide.