A German consortium released Soofi S 30B-A3B, an open model that beats OLMo 3 32B and Apertus 70B on both English and German benchmarks. The 31.6B-parameter mixture-of-experts model activates only 3.2B parameters per token and was trained entirely on Deutsche Telekom's Industrial AI Cloud in Munich.
Key facts
- 31.6B total params, 3.2B active per token.
- 27 trillion tokens processed across three training phases.
- 8x tokens/s/GPU vs dense 14-24B models at 40K context.
- Throughput flat from 4K to 256K tokens.
- Trained on Deutsche Telekom's Munich AI cloud.
Architecture and efficiency gains
Soofi S adopts Nvidia's Nemotron 3 Nano architecture without modification, combining Mamba-2 layers with standard attention layers. Only 6 of the model's 52 layers maintain a KV cache, which eliminates the memory bottleneck that slows conventional transformers at long context lengths. At 40,000-token context with 32 parallel requests, Soofi S generates roughly eight times more tokens per second per GPU than dense models in the 14–24 billion parameter range according to The Decoder. Throughput stays nearly flat from 4,000 to 256,000 tokens, a behavior shared only by Alibaba's Qwen3.5 35B-A3B among measured models.
Training data and benchmark results
The consortium processed about 27 trillion tokens across three phases. The first phase used roughly 20 trillion tokens from a broad mix of web, code, math, and domain-specific texts. A second phase of about 6 trillion tokens came from higher-quality sources, and a third shorter phase extended the context window by training on very long documents up to 256,000 tokens. The training mix was deliberately weighted toward German text, which the consortium says explains Soofi S's top scores on German benchmarks. The model also leads on English and programming tasks among fully open models.
Significance for European AI sovereignty
Soofi S is one of the first large language models trained entirely on Deutsche Telekom's Industrial AI Cloud in Munich, marking a milestone for European AI infrastructure independence. The model is released as open-source under an unspecified license—the consortium did not disclose the exact license terms. The KI Bundesverband (German AI Association) coordinated the project, which involved multiple research institutions across Germany.
What to watch
Watch for downstream fine-tuned variants on Hugging Face over the next 90 days, and whether European enterprises adopt Soofi S for German-language applications. Also track if the consortium releases benchmark numbers comparing against closed models like GPT-4o or Claude 3.5, which were not included in the reported comparisons.
Source: the-decoder.com
