30B total / ~3B active (hybrid MoE; Hugging Face cards also cite ~3.5B active)
Unknown
Custom / Other (paper says highly permissive open-access terms; Hugging Face cards still show placeholder license metadata)
Official first-generation Soofi release family from a German research consortium, positioned as a sovereign European open-source alternative to US and Chinese foundation models for industrial deployment, code generation, and agentic enterprise use. The public paper and model cards describe Soofi S as a custom hybrid Mamba-2 and Transformer Mixture-of-Experts model trained from scratch in German sovereign infrastructure, with 30B total parameters, roughly 3B active parameters per token, 128 routing experts, 23 Mamba-2/MoE layers plus 6 attention layers, and near-constant inference-cache growth that is framed as a throughput advantage over dense long-context models. The paper says the model was pretrained on about 27 trillion tokens with deliberate German up-weighting and built end-to-end on Deutsche Telekom’s Industrial AI Cloud in Munich. Public release materials position the base checkpoint as a research and post-training foundation, while the companion instruction and reasoning variants add assistant behavior, native tool-calling support, and explicit thinking traces for the two reasoning previews, Isar and Rhine. The Soofi paper claims the base model matches dense 14B to 27B models on aggregate English and German benchmarks, achieves the best code aggregates in both languages among 17 open base models, and outperforms other European sovereign baselines including models with far more active parameters. The current Hugging Face release is still explicitly labeled as preview or internal checkpoint material whose weights and metadata may change, and the model cards currently use a custom 'Other' license placeholder while the paper describes the intended release terms as highly permissive open-access terms with weights, intermediate checkpoints, hyperparameters, and training and evaluation code.