Soofi S 30B-A3B Tops Olmo 3 and Apertus in Open-Model Benchmarks

A group calling itself the Soofi-Team has posted a paper on arXiv describing Soofi S 30B-A3B, a Mixture-of-Experts foundation model for German and English that they say leads every fully open model on aggregate scores in both languages.

The architecture is a hybrid Mamba Transformer that activates only 3B of 30B parameters per token, pretrained on roughly 27 trillion tokens with German deliberately up-weighted. The hybrid design keeps the inference cache near-constant as context grows, which the authors argue gives it a decisive throughput advantage over dense models for long-context, high-concurrency deployment. Their headline benchmark claim is that Soofi S matches dense 14 to 27B models on English and German aggregates, wins the code aggregates in both languages among 17 open base models they compared, and beats every European sovereign baseline in that comparison, including ones far larger in active parameters. Among the fully open models, they place Soofi S ahead of Olmo 3 32B and Apertus 70B.

The sovereignty piece is not just a slogan. The model was built end-to-end on the German Industrial AI Cloud, a sovereign HPC-scale AI infrastructure operated by Deutsche Telekom in Munich. That matters for German enterprises and public-sector buyers who have been reluctant to commit to US-hosted frontier labs for regulatory or procurement reasons. The release plan is unusually thorough for a project this size: weights, selected intermediate checkpoints, full per-source data accounting, hyperparameters, and training and evaluation code, all under permissive terms where the underlying licenses allow.

The honest caveat is that these leaderboard claims come from the authors' own comparison rather than an independent evaluation, and the comparison set is limited to open models. Commercially licensed training sources are only documented with aggregate statistics and exact mixture accounting, which downstream users will want to inspect carefully before shipping on the weights. What the paper does not give you is head-to-head performance against the closed US frontier models most German buyers are actually choosing between.

If the numbers survive replication, the practical shift is real: a permissively licensed German-first base model, trained on sovereign infrastructure, becomes something a German bank or ministry can plausibly build on without an American dependency.