German Consortium Releases Soofi S, a 30B Open-Weights Model Tuned for German and English
A publicly funded German consortium released Soofi S, a 30B-parameter hybrid Mamba-Transformer model that tops fully open-weights rivals on English and German benchmarks, including OLMo 3 32B and Apertus 70B.
The Soofi consortium, led by Germany's KI Bundesverband with Fraunhofer institutes, DFKI, and companies including Ellamind and Merantix Momentum, released Soofi S 30B-A3B on July 13, an open-weights model trained on Deutsche Telekom's Industrial AI Cloud in Munich.
- 31.6B total parameters, 3.2B active per token, hybrid Mamba-2/attention architecture (only 6 of 52 layers keep a KV cache)
- Trained on ~27 trillion tokens using up to 512 Nvidia B200 GPUs, about 253,000 GPU-hours
- English aggregate score 70.1, German aggregate 79.1, HumanEval 73.8%, MBPP 70.2%
- Beats OLMo 3 32B and Apertus 70B across categories among fully open models
- Funded with roughly EUR 20 million by Germany's economics ministry under the EU's IPCEI-CIS program
Weights, intermediate checkpoints, training code, and a full data inventory are on Hugging Face. The consortium says 99% of the training mix is independently reconstructible, meeting the Open Source AI Definition 1.0, though the final license wasn't set at release. Try comparable open-weights models in the coding arena.