Soofi S 30B-A3B — The German Open AI Model That Tops Benchmarks in Both English and German

Soofi S — A Milestone for European AI Sovereignty

On July 13, 2026, a German research consortium coordinated by KI Bundesverband (German AI Association) officially released Soofi S 30B-A3B — an open-source language model trained entirely on the Deutsche Telekom industrial AI cloud in Munich. This is one of the first major language models built entirely on German sovereign infrastructure.

The model achieved impressive results: in both English and German benchmarks, Soofi S outperforms all other fully open models, including OLMo 3 32B (Allen Institute for AI) and Apertus 70B (ETH Zurich / EPFL). What makes this particularly notable is that Soofi S achieves these results with far fewer active parameters — just 3.2 billion per token.

This article offers a detailed look at Soofi S's architecture, training methodology, benchmark results, and its significance for European AI sovereignty.

Hybrid Mamba-Transformer Architecture: Efficiency First

One of Soofi S's most interesting features is its architecture. The model uses Mixture-of-Experts (MoE) — it has 31.6 billion parameters in total, but activates only 3.2 billion for each generated token. This means computational cost is closer to a 3B model than a traditional 30B model.

The consortium used Nvidia's Nemotron 3 Nano architecture without modifications — a hybrid design that combines Mamba-2 layers with standard attention layers. Mamba layers, from the State Space Models (SSM) family, enable processing of long sequences in linear time (O(n) vs O(n²) for traditional Transformers).

This hybrid approach is particularly effective: at 40K context length, Soofi S's Mamba component delivers 8x higher throughput than an equivalent dense model, without quality degradation.

Training: 27 Trillion Tokens with German Emphasis

Soofi S was trained on 27 trillion tokens, comparable to Meta's Llama 3.1 70B's 15T tokens, but with significantly more German language data. The training proceeded in two phases:

In the first phase (22T tokens), 7.2% of the data was German. In the second phase (5T tokens), the German share increased to 15.3%. In total, 8.7% of Soofi S's training data is German — far more than any other open model except the German-specific Leo models.

Training took place on 512 Nvidia B200 GPUs, all hosted on Deutsche Telekom's Munich industrial AI cloud. This means the entire data flow and computation stayed within Germany — important for GDPR compliance.

Benchmarks: Leading in Both English and German

Soofi S delivered impressive results across multiple benchmarks. In English, it outperformed OLMo 3 32B on tests including MMLU-Pro (massive multi-task language understanding), DROP (text reasoning), and ARC-Challenge (scientific reasoning).

In German, Soofi S achieved the highest scores on HellaSwag-DE, ARC-DE, MMLU-DE, and DROP-DE benchmarks, surpassing OLMo 3 32B, Apertus 70B, Leo-Mistral-7B, Leo-Scarlet-13B, Llama 3.1 70B, Qwen 2.5 32B, Falcon 3, Gemma 2, and Llama 4 Scout.

Notably, Soofi S outperformed Llama 4 Scout (109B) in German benchmarks despite having 30x fewer active parameters. This confirms that targeted training data is often more important than sheer model size.

In code generation, Soofi S achieved scores close to OLMo 3 32B, though this area still needs development — on EnglishCodeDS-1000, Soofi S scored 67.5% versus OLMo 3 32B's 72%.

OSI Open Source AI Definition: Truly Open

Soofi S fully complies with the OSI (Open Source Initiative) Open Source AI Definition 1.0. This means model weights are fully open, training datasets are available without restrictions, training code and methodology are published, and any third party can download, modify, and use the model commercially.

This is significant because many "open" models (e.g., Llama, Qwen) actually use restricted licenses that limit commercial use. Soofi S is different — it is genuinely open.

European AI Sovereignty

One of Soofi S's most important aspects is its geopolitical context. Europe has long sought to build its own AI infrastructure not dependent on American or Chinese platforms. Soofi S represents one of the most successful examples of this effort.

Germany, as Europe's largest economy, has been particularly active in AI. The KI Bundesverband brings together over 400 members from research institutes, universities, and industry. Soofi S's success demonstrates that Europe can compete with global AI leaders.

What This Means for Georgia

Soofi S's release has direct implications for the Georgian AI ecosystem. Open models like Soofi S mean Georgian startups and research groups can use world-class AI infrastructure to build their own products.

Soofi S's German language support is particularly interesting — it shows that high-quality language models can be built for non-English languages too. This is an encouraging sign for Georgian language AI integration prospects.

At SiTech, we closely follow the development of open models. Soofi S's success confirms that the future of AI is open, collaborative, and multilingual.

Conclusion

Soofi S 30B-A3B represents a significant milestone in the development of open AI models. It is not only technically impressive — hybrid Mamba-Transformer architecture, MoE, 27T tokens — but also symbolic as a demonstration of European AI sovereignty.

The model's openness, high performance, and German language focus make it attractive for both research and commercial use. We look forward to seeing how this direction develops.