Soofi S: German AI consortium unveils open 30B model leading English and German benchmarks

A German research consortium has released Soofi S, an open-source language model trained entirely on Deutsche Telekom’s AI cloud infrastructure in Munich. Jonathan Kemper reports for The Decoder that the 30B model tops benchmarks in both English and German among fully open models, surpassing OLMo 3 32B and Apertus 70B.

Efficient by design

Soofi S uses a mixture-of-experts architecture based on Nvidia’s Nemotron 3 Nano, activating only 3.2 of its 31.6 billion parameters per token. Combined with a hybrid Mamba-Transformer design, this keeps generation speed nearly constant even at context lengths up to 256,000 tokens, unlike dense competitors that slow down sharply.

The consortium, coordinated by the KI Bundesverband, trained the model on about 27 trillion tokens with a deliberate focus on German sources, including web text, newspaper archives, and synthetic data. German makes up up to 15.3 percent of later training phases, far above typical international models.

The model leads on coding benchmarks and German-language tests but lags on competition math and factual retrieval, likely due to its small active parameter count.

Overtraining dispute

Critics argued Soofi S was trained on far more tokens than classic Chinchilla scaling laws recommend. Technical lead Michael Fromm disputes this, arguing those rules don’t apply to MoE architectures, since individual experts benefit from repeated exposure to high-quality data. He points to Nvidia’s own models, trained on up to 25 trillion tokens, as precedent.

The consortium has released model weights, training code, and a detailed data inventory, meeting the Open Source AI Definition 1.0, though a stricter European open-data standard remains unmet due to a small share of commercially licensed data.