Soofi S 30B activates 3B parameters per token, tops European AI baselines
Deutsche Telekom's Soofi-Team released Soofi S, a 30B-parameter Mixture-of-Experts foundation model that activates only 3B parameters per token and outperforms European baselines on German and English benchmarks.
Soofi S 30B-A3B, a Mixture-of-Experts foundation model from Deutsche Telekom's Soofi-Team, activates only 3 billion of its 30 billion parameters per token while maintaining near-constant inference cache as context grows. The hybrid Mamba Transformer was pretrained on roughly 27 trillion tokens with deliberately up-weighted German content, trained end-to-end on the German Industrial AI Cloud—a sovereign HPC-scale infrastructure operated by Deutsche Telekom in Munich. The sparse-activation design gives Soofi S a decisive throughput advantage over dense models for long-context, high-concurrency deployment: by activating only 3B parameters per forward pass instead of the full 30B, the model keeps memory footprint and compute costs closer to a small dense model while retaining the representational capacity of a much larger one.
On aggregate English and German benchmarks, Soofi S matches dense models in the 14B to 27B range while achieving the best code performance in both languages among 17 open base models tested. Among fully open models, it obtains the highest English and German evaluation scores, ahead of Olmo 3 32B and Apertus 70B, and outperforms every European sovereign baseline in the team's comparison, including ones with far more active parameters. The release includes weights, selected intermediate checkpoints, full per-source data accounting, hyperparameters, and training and evaluation code under highly permissive open-access terms; where source licenses permit, data-construction artifacts are released under permissive licenses, while commercially licensed sources are documented with aggregate statistics and exact mixture accounting. The preprint appeared on July 13, 2026.
