German AI consortium releases Soofi S, an open 30B model that tops benchmarks in both English and German
Summary
The Decoder reports on Soofi S, a German open-source 30B language model that allegedly tops benchmarks in English and German. The model uses a lean mixture-of-experts hybrid architecture to maintain throughput with long contexts, trained on a German-heavy data mix across three phases, and released with weights, intermediate checkpoints, and a detailed data inventory. The piece also covers training infrastructure, governance, and ongoing debate about training data and scaling laws, noting that results are self-reported and subject to verification.