Soofi S was built end-to-end on the German Industrial AI Cloud, a sovereign HPC-scale AI infrastructure operated by Deutsche Telekom in Munich. That's not a marketing claim — it means the training compute, the data, and the infrastructure are all inside German/EU legal jurisdiction, with no dependency on AWS, Azure, GCP, or any US cloud provider. For GDPR compliance, for government procurement, for sectors with data residency requirements, this matters in a way that "open weights but trained on Nvidia's US cluster" doesn't.
The architecture is genuinely novel
The hybrid design activates only 3B of 30B parameters per token and keeps the inference cache near-constant as context grows — that "near-constant cache" claim is the key technical differentiator. Standard Transformer attention has a KV cache that grows linearly with context length. The hybrid Mamba-2/MoE architecture uses 23 Mamba-2/MoE layers and 6 attention layers, with 128 routing experts and 6 active per token. Mamba is a state-space model — it maintains a fixed-size recurrent state rather than caching all previous tokens. The result is that very long contexts don't cost proportionally more memory. For industrial document processing and long-context enterprise use cases, that's a real advantage over pure Transformer architectures.
The transparency claim is the strongest in the database
Soofi S will be released under highly permissive, open-access terms: weights, selected intermediate checkpoints, full per-source data accounting, hyperparameters, and training and evaluation code. Where source licenses permit, data-construction artifacts are released under permissive licenses; commercially licensed sources are documented with aggregate statistics and exact mixture accounting.
No other model in the Frontier Arcade database has published full per-source data accounting. Most labs won't even name their data sources. Soofi S is releasing the exact mixture — every source, what percentage of training tokens came from it, what license it carries. That's a level of transparency the field has never seen at this scale.
Rivers, not just release names
The Rhine name (Soofi-S-Rhine-Preview, one of the two reasoning variants) is a river. The Isar name (the other reasoning variant) is also a river — Munich's river specifically. The naming is a deliberate European geography signal.
Links
- Paper: A Sovereign, Open-Source Foundation Model for German and English (arXiv:2607.09424)
- Model collection: Soofi-S Beta Models on Hugging Face
- Base model: Soofi-Project/Soofi-S-Base
- Reasoning preview: Soofi-Project/Soofi-S-Rhine-Preview