German research consortium releases Soofi-S, an open 30B sovereign AI model for German and English

A German research consortium has released Soofi-S-30B-A3B, an open-weights foundation model built as a sovereign European alternative to US and Chinese AI systems, with training run entirely on European cloud infrastructure.

What's new

The model card states: "Soofi-S-30B-A3B is a pre-trained base model that generates free-form text continuations for a given prompt. It was developed by a German research consortium."

The project name, SOOFI, stands for Sovereign Open Source Foundation Models — a description of its explicit mission. Key specifications:

• Architecture: a hybrid Mixture-of-Experts design combining Mamba-2 state-space layers with standard attention layers — 23 Mamba-2/MoE layers plus 6 attention layers, with 128 routing experts
• Parameters: 30 billion total, with roughly 3.5 billion active per token during inference
• Training data: built from scratch on approximately 25 trillion tokens, with English and German as the primary languages
• Training infrastructure: hosted entirely in Europe, on T-Systems' (Deutsche Telekom) Industrial AI Cloud, to keep data and compute inside European jurisdiction
• Status: released as a base model — it has not been instruction-tuned, aligned, or safety-tuned, and is intended as a foundation for further fine-tuning (SFT, preference tuning, domain adaptation) rather than direct deployment as an assistant

The model requires trust_remote_code=True to load with the transformers library, since it ships with custom modeling code for its hybrid architecture.

Context

Soofi-S adopts the Mamba-2/attention hybrid architecture pioneered by NVIDIA's Nemotron 3 Nano rather than a novel design of its own, applying it to a specifically German/English-focused pretraining run. The consortium coordinating the project draws on Germany's applied-research infrastructure — the effort was funded in part through Germany's Federal Ministry for Economic Affairs and Energy under the European IPCEI-CIS program, with participation from institutions including Fraunhofer IAIS, Fraunhofer IIS, DFKI, TU Darmstadt, and the University of Würzburg, among others.

The release fits a broader pattern of European sovereign-AI initiatives — government- and consortium-backed efforts to build foundation models that don't depend on US labs (OpenAI, Anthropic, Google, Meta) or Chinese labs (DeepSeek, Qwen, Kimi) for critical infrastructure, particularly for languages and use cases underserved by general-purpose frontier models trained primarily on English data.

Why it matters

Sovereignty is the explicit design goal here, not a side effect: training on Deutsche Telekom's own European cloud, rather than US hyperscaler infrastructure, means the entire pipeline — from raw compute to trained weights — stays under European jurisdiction. For government and industrial users in Germany and the EU who face regulatory or contractual restrictions on where AI training and inference can happen, that matters more than raw benchmark performance.

Releasing as an open base model, rather than a hosted API, also means downstream users can fine-tune, audit, and deploy Soofi-S entirely on their own infrastructure — a meaningful difference for defense, industrial, and public-sector use cases where dependency on a foreign API provider is itself treated as a risk. Choosing NVIDIA's Nemotron 3 Nano hybrid architecture rather than inventing a new one also signals a pragmatic strategy: European sovereignty efforts increasingly compete on data, training infrastructure, and governance rather than trying to out-architect the frontier labs on model design.