Soofi S: a German public consortium ships a competitive open model

A German research consortium has released Soofi S, an open model that shows a European public effort can match the open models coming out of larger labs. The project was coordinated by the KI Bundesverband and funded by Germany's Federal Ministry for Economic Affairs and Energy, with work from Fraunhofer institutes, DFKI, TU Darmstadt, and companies including Ellamind and Merantix Momentum. It is a mixture-of-experts model with 31.6 billion total parameters that activates only 3.2 billion per token, so it stays cheap to run for its size.

The design is unusual. Soofi S uses a hybrid Mamba-Transformer architecture in which only 6 of its 52 layers keep a KV cache, which is what lets it handle context up to 256,000 tokens without the usual memory cost. On benchmarks it beats other open models of similar or larger size, including OLMo 3 32B and Apertus 70B, scoring 70.1 on English aggregates, 79.1 on German ones, and 73.8% on HumanEval for code.

The training run is documented in a way most releases avoid. The team trained on roughly 27 trillion tokens across three phases between March and May 2026, using up to 512 Nvidia B200 GPUs at a Deutsche Telekom site in Munich, for about 253,000 GPU-hours. German data rose from 7.2% of the mix in the first phase to 15.3% in the second. One caveat sits in the license: the model meets the Open Source AI Definition 1.0, but 1.3% of the data comes from a commercial source (Genios), which keeps it from meeting stricter European open-data rules.

Why it matters

If you work in Europe and need a capable open model you can host yourself, Soofi S gives you strong German performance and a documented training recipe, though the small slice of commercial data means you should check the license before shipping it in a regulated product.