Researchers have introduced Soofi S 30B-A3B, a new open-source foundation model designed for German and English languages. This model utilizes a hybrid Mamba-Transformer architecture, activating only 3 billion of its 30 billion parameters per token, which offers significant throughput advantages for long-context applications. Soofi S was trained on approximately 27 trillion tokens and has demonstrated competitive performance against larger dense models, particularly excelling in code generation tasks and outperforming other European sovereign models. The development was conducted on Deutsche Telekom's German Industrial AI Cloud infrastructure in Munich, and the model will be released with permissive open-access terms, including weights, code, and detailed data accounting. AI
IMPACT This open-source model could accelerate development and deployment of German and English language AI applications, particularly those requiring long contexts and high concurrency.
RANK_REASON The cluster describes a new research paper detailing an open-source foundation model.
Read on Hugging Face Daily Papers →
- Apertus-70B
- Deutsche Telekom
- English
- German
- German Industrial AI Cloud
- Hugging Face
- Mamba-Transformer
- Munich
- Olmo-3-32B
- Soofi S 30B-A3B
AI-generated summary · Google Gemini · from 3 sources. How we write summaries →