Soofi S Surfaces on Hacker News: Germany's Sovereign 30B-A3B MoE, 27T Tokens, Trained Entirely on Deutsche Telekom Infrastructure

Hacker News / arXiv:2607.09424 (Fraunhofer IIS press release)·medium signal

Soofi ('Sovereign Open Source Foundation Models'), an IPCEI-CIS/8ra-funded German consortium project, hit the Hacker News front page (51 points) with Soofi S — a 30B Mixture-of-Experts activating only 3B parameters per token, pretrained on ~27 trillion tokens with deliberate German-language weighting. The paper (arXiv:2607.09424) claims it matches dense 14–27B models on aggregate English and German benchmarks and posts the best code aggregates in both languages among 17 open base models, with near-constant inference cache as context grows. It was trained entirely on Deutsche Telekom's German Industrial AI Cloud in Munich; the team plans to release weights, intermediate checkpoints, training code, and full data accounting, though general availability hasn't landed yet.