Soofi S beats OLMo on German benchmarks at 3.2B active params

a german research consortium trained the 30b open-source model soofi s entirely on deutsche telekom's ai cloud, using a hybrid architecture that activates just 3.2b parameters per token to match the compute cost of much smaller models. by focusing 15% of phase 2 training data on german sources, the model surpasses fully open competitors like olmo 3 32b and apertus 70b on german and english benchmarks. its design avoids the throughput drop-off typical of dense transformers, delivering stable generation speeds from 4k to 256k tokens. this release provides a high-performance, resource-efficient option for german-language applications that requires less hardware than comparable dense models.