German Consortium Unveils Soofi S: Open 30B Model Tops Benchmarks
German AI consortium releases Soofi S, an open 30B model that tops benchmarks
I am excited to share that the German AI consortium has released Soofi S, an open 30B model trained on Deutsche Telekom's Industrial AI Cloud. This hybrid Mamba-Transformer architecture outperforms competitors like OLMo 3 and Apertus 70B in both English and German benchmarks. Despite criticism regarding its training data volume, the team argues that Mixture-of-Experts models defy traditional Chinchilla scaling laws, delivering superior efficiency and long-context performance.
There's new research showing that the old scaling laws from dense models no longer apply to MoE architectures.
- jbellis
It tops benchmarks because it uses them in its training data. https://x.com/eliebakouch/status/2077425801633427919
- skeledrew
> the facility runs entirely on renewable energy, is cooled with water from the Eisbach canal, and feeds waste heat into the surrounding Tucherpark neighborhood.
That's it right there. Show the others how it's done. Can see going forward most model training being done in winter, making for a proper resource recycling ecosystem.
- mft_
Great to see more competition in this space, and especially from Europe, but... it's a shame when the benchmarks don't include the current best comparable models. They shows results against Qwen 3.5 and Gemma 3, but Qwen 3.6 and Gemma 4 have been available for ~three months.
- throwa356262
I think this news is more about the infrastructure than the model.
Either way, happy to see two open models from outside the us-cn duopoly in the same week.
- beklein
Some other (as in better) sources I found:
- https://huggingface.co/spaces/Soofi-Project/Pretraining-Tech...