German AI model Soofi S tops open benchmarks in English and German

Soofi S represents a deliberate effort by European institutions to build competitive open-source AI infrastructure independent of non-European providers. The consortium—coordinated by the German AI Association and funded by the German Federal Ministry for Economic Affairs and Energy as part of the European IPCEI-CIS program—includes the Fraunhofer Institutes IAIS and IIS, the German Research Center for Artificial Intelligence (DFKI), TU Darmstadt, the University of Würzburg, the L3S Research Center, the Berlin University of Applied Sciences, and companies Ellamind and Merantix Momentum. The model's training on sovereign infrastructure in Munich, powered entirely by renewable energy, signals a commitment to reducing dependence on externally controlled cloud providers.

The hybrid architecture choice has concrete practical implications. By combining Mamba-2 layers (which have superior memory efficiency) with selective use of standard attention, Soofi S avoids the linear memory growth that makes long-context inference expensive for conventional transformers. The benchmark results bear this out: while models like Apertus 70B and Qwen3 32B see sharp throughput drops as context length increases, Soofi S maintains nearly flat performance from 4,000 to 256,000 tokens. This efficiency appears valuable for industrial applications involving lengthy technical documents—exactly the use case the consortium now plans to validate with partners.

The training data composition reveals a strategic trade-off. By weighting German-language content far more heavily than Nvidia's reference recipe (15.3 percent versus 5 percent for all non-English languages combined), the consortium achieved top-tier German benchmark performance without sacrificing English capability. However, the model does show specific weaknesses: it scores significantly lower than Qwen3.5 on German math benchmarks (56 points on Minerva MATH-DE versus 76.5 for Qwen3.5) and exhibits a notable failure mode on the RULER extraction task at very long contexts (3 percent hit rate beyond 32,000 tokens versus 60–64 percent for Nemotron). The authors attribute the extraction weakness to a lack of synthetic long-context extraction data in the training set, identifying a concrete direction for future improvement. Regarding openness, Soofi S meets the Open Source Initiative's Open Source AI Definition 1.0, though it does not satisfy a stricter European standard that would require all training tokens to be freely distributable; the 1.3 percent share of commercially licensed Genios newspaper data prevents full alignment with that stricter threshold.