Germany Takes on US With Strongest Non-Chinese Open-Source AI Model

A German research consortium has released Soofi S, a new open-source large language model designed for German and English.

The model, officially called Soofi S 30B-A3B, is described as a sovereign, open-source foundation model built for Europe’s AI ecosystem. According to the project’s technical report, it is a Mixture-of-Experts hybrid Mamba-Transformer model trained primarily for German and English tasks.

What Is Soofi S?

Soofi S is designed as an open alternative to large AI models from the US and China.

The model has about 31.6 billion total parameters, but activates only around 3.2 billion per generated token. This means it can offer the scale of a larger model while keeping compute requirements closer to a much smaller system.

The project says the model was trained from scratch on about 27 trillion tokens, with German data deliberately given more weight during training. It was built end-to-end on Deutsche Telekom’s Industrial AI Cloud in Munich.

Mixture-of-Experts Architecture

Soofi S uses a Mixture-of-Experts, or MoE, architecture.

In simple terms, MoE divides an AI model into smaller expert networks. Only some of these experts are used for each token or task, which helps reduce computing cost while still allowing the model to benefit from a larger overall structure.

This architecture has become common in leading open-AI models. Nvidia, citing Artificial Analysis data, said the top 10 most intelligent open-source models all use Mixture-of-Experts designs.

Benchmark Claims

The Soofi team says Soofi S performs strongly against open-source competitors.

According to the model’s technical report, Soofi S achieved the highest English and German evaluation scores among fully open models tested, ahead of models such as Olmo 3 32B and Apertus 70B. It also recorded the best code aggregate scores in both English and German among 17 open base models tested by the team.

The model also outperformed European sovereign AI baselines included in the comparison, including models with more active parameters.

Built for Sovereign AI

The consortium is positioning Soofi S as part of Europe’s push for sovereign AI.

The model is intended to give companies, researchers, and public-sector users an open-source option that can be hosted and developed within European infrastructure. Its Hugging Face page says the project is designed to provide a secure European open-source alternative for industrial use.

The Soofi consortium includes German research and industry partners, with coordination by KI Bundesverband and participation from groups such as Fraunhofer IAIS, DFKI, TU Darmstadt, ellamind, and Merantix Momentum.

Open Release

The Soofi team says it plans to release the model under permissive open-access terms.

This includes model weights, selected intermediate checkpoints, training details, evaluation code, hyperparameters, and data accounting where licensing allows.

For now, Soofi S adds another major European entry to the open AI model race, with a focus on German-language performance, English capability, coding benchmarks, and sovereign infrastructure.