arXiv:2607.09424: Soofi S — Sovereign Open-Source Model for German and English Trained on Telekom's Cloud

The research team (Soofi-Team) has published a paper on arXiv about Soofi S — a sovereign, fully open language model for German and English. The model was pre-trained exclusively on Deutsche Telekom’s cloud infrastructure in Munich, as part of the German Industrial AI Cloud initiative, placing it in the rare category of European AI models with verified data sovereignty.

What Is Sovereign AI and Why Is the Architecture Hybrid?

A sovereign AI model refers to a system that is trained and hosted exclusively on domestic or regional infrastructure, without dependence on foreign computing resource providers — allowing organizations to retain full control over their data and processes. Soofi S uses a hybrid Mamba-Transformer architecture in a Mixture-of-Experts (MoE) configuration: a total of 30 billion parameters, of which only 3 billion are active in each pass. This ratio — 10:1 between total and active parameters — is typical for MoE models and means Soofi S achieves capabilities comparable to dense models of 14 to 27 billion parameters with significantly lower inference computational cost.

How Much Data and What Are the Results?

The model was pre-trained on approximately 27 trillion tokens with deliberate augmentation of German language data — unusual since most global models treat German as a minority language. On aggregated EN/DE benchmarks Soofi S is on par with dense models of 14 to 27B, and it takes first place on the code benchmark among 17 open baseline models tested in both languages simultaneously. Compared to OLMo 3 32B and Apertus 70B — considered the leading fully open European alternatives — Soofi S outperforms both, despite having fewer active parameters.

Openness and Reproducibility

The authors publish fully open weights, checkpoints, hyperparameters, and evaluation code — a level of transparency rare even among models that call themselves “open-source.” This enables researchers and companies to fine-tune on their own data without dependence on a central service provider. The preprint is available at arXiv:2607.09424.