The paper presents Soofi S 30B-A3B, a German-English sovereign foundation model combining Mixture-of-Experts routing with a hybrid Mamba-Transformer architecture. It contains 30B total parameters but activates about 3B per token, while its inference cache is designed to remain nearly constant as context length grows. Pretrained on roughly 27T tokens with German up-weighted, the authors report performance comparable to dense 14B–27B models, leading code aggregates across 17 open base models, and stronger results than the European sovereign baselines evaluated. Training was conducted on Deutsche Telekom’s German Industrial AI Cloud in Munich.
No heat snapshots are available in the last 24 hours.