Germany Unveils Soofi S, a Fully Open AI Model Aiming to Be Europe’s Answer to OpenAI and Qwen

Germany has hit a new milestone in European artificial intelligence. This week, the Soofi consortium unveiled Soofi S, a 30-billion-parameter language model built entirely in Europe. According to its creators, it’s currently the most powerful fully open AI model in the world—Europe’s proof that it can build cutting-edge AI without U.S. or Chinese tech.

The launch comes as the European Union doubles down on digital sovereignty. While the market is dominated by closed models from OpenAI, Anthropic, and Google—and Chinese players like Alibaba’s Qwen are rapidly gaining ground—Soofi S leans hard into maximum transparency.

You can find the research on the model here.

Soofi S is the result of a collaboration between German universities, research institutes, and AI companies led by the German AI Association (KI Bundesverband). It’s a 30-billion-parameter Mixture-of-Experts (MoE) model. Thanks to that architecture, only about 3 billion parameters are active during inference, making it far more efficient than traditional models of similar size.

But performance isn’t the only differentiator.

The consortium is releasing not just the model weights, but also the training code, documentation, and detailed dataset information. That means Soofi S meets the strictest definition of open source—something only a handful of large AI models can claim.

The developers are strikingly candid about the model’s capabilities.

Michael Fromm, one of the project’s technical leads, writes: “Soofi S is the most powerful fully open model. Qwen 3.5 leads in pure capacity; our strengths are capacity, throughput, and full openness.”

In other words, Alibaba’s Qwen 3.5 still tops some AI benchmarks. But the consortium argues Soofi S delivers a stronger mix of performance, speed, and total transparency.

That makes it especially appealing to companies, governments, and researchers who want full control over their AI infrastructure.

Soofi S was trained entirely on Deutsche Telekom’s German Industrial AI Cloud in Munich. Training used roughly 27 trillion tokens, with a strong focus on German- and English-language data.

The project aligns closely with Europe’s push to reduce reliance on U.S. cloud providers and Chinese AI models. For industry, government, healthcare, and research, a fully open model offers crucial transparency and control.

The debut of Soofi S signals that Europe is stepping up in the global AI race. While the model isn’t number one on every benchmark, it proves European organizations can build world-class foundation models—and release them fully open.

In a market where transparency is increasingly non-negotiable, that could be a major advantage. Soofi S lets developers and organizations run, customize, and audit a state-of-the-art language model on their own terms—without dependence on closed commercial platforms.

For Europe, it’s more than a technical feat. It’s a strategic statement that the region is serious about building its own independent AI infrastructure.