Skip to content

Soofi S: German Open-Source Language Model with Hybrid Architecture

In brief: Soofi S 30B-A3B is an open-source language model trained on German data with a mixture-of-experts architecture that provides throughput advantages for long contexts through selective parameter activation (3B/30B).

Researchers have developed Soofi S 30B-A3B, a sovereign open-source language model for German and English trained on European HPC infrastructure. The model combines Mamba and Transformer architecture in a mixture-of-experts design and activates only 3 billion of 30 billion parameters per token.

Soofi S 30B-A3B is a mixture-of-experts hybrid architecture combining Mamba and Transformer components, pretrained on approximately 27 trillion tokens with targeted overweighting of German training data. The model activates only 3 billion of 30 billion total parameters per token and maintains a constant inference cache size as context grows.

Soofi S achieves performance comparable to dense models between 14 and 27 billion parameters on English and German benchmarks and achieves the best code aggregates among 17 open base models in both languages. Among fully open models, Soofi S outperforms competitors like Olmo 3 32B and Apertus 70B for English and German according to evaluation. All evaluated European sovereignty baselines are exceeded.

For CTOs, the throughput advantage in long-context and high-concurrency scenarios is relevant: through activated 3B instead of 30B parameters and constant cache size, latency and resource overhead are significantly reduced compared to dense models. The model was trained entirely on the German Industrial AI Cloud, an HPC infrastructure operated by Deutsche Telekom in Munich.

Soofi S is released under highly permissive open-access conditions: weights, selected intermediate checkpoints, complete data provenance documentation, hyperparameters, and training and evaluation code are released. For permissive source licenses, data construction artifacts are also published; commercially licensed sources are documented with aggregate statistics and exact mixture accounting.


Source: arxiv.org · Published 9 July 2026
Lumi AI News — AI-assisted curation according to Art. 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.

Share on: