FULL STORY

Inside Soofi S: Germany's Open-Source Sovereign AI Model

A German consortium launched Soofi S, an open-source bilingual model championing sovereign AI, and subsequently released its technical report detailing a 30B hybrid Mamba architecture.

2026-07-12 ~ 2026-07-23 · 3 episodes · 9 posts

Episode 1 · German Alliance Releases Open-Source Sovereign Model Soofi S 30B-A3B (2026-07-12, 5 posts)

A coalition of German research institutions and companies released Soofi S 30B-A3B, an open-source foundation model for German and English. Marketed as a "sovereign model," the release has garnered community attention for its extreme transparency, providing not just model weights but a comprehensive set of training artifacts including data splits, hyperparameters, and evaluation code.

Key Details

According to @The Soofi-Team, Soofi S 30B-A3B utilizes a MoE and Hybrid Mamba Transformer architecture, with a total of 30B parameters but only 3B active parameters per token. @mishig25 noted that the project involves institutions like DFKI, Fraunhofer IAIS / IIS, TU Darmstadt, and others. @eliebakouch highlighted that the release includes full training reports, data source splits, hyperparameters, and checkpoints, achieving strong performance in English and German evaluations. Regarding training scale, posts mention both 27T tokens and approximately 26T tokens.

Controversies and Questions

Regarding the technical novelty of this "sovereign model," some community members expressed cautious evaluations. @eliebakouch pointed out in another comment that the model's architecture and most hyperparameters largely follow Nemotron 3 Nano, with the primary change being an approximate 20% adjustment in the data mixture ratio. This sparked discussions on whether the model's innovation stems from architectural breakthroughs or merely data ratio adjustments.

Episode 2 · German AI Consortium Releases Open-Source Bilingual Model Soofi S (2026-07-15, 2 posts)

The German Soofi consortium released Soofi S, a 30B parameter open-source model trained in Europe. It achieves leading bilingual performance in English and German benchmarks.

Episode 3 · Soofi S Releases 30B Hybrid Mamba Open-Source Model (2026-07-21, 2 posts)

Soofi S released the full pretraining report for its 30B-A3B MoE hybrid Mamba model. Trained on 27 trillion tokens, the open-source model claims leading performance in German-English bilingual evaluations.