A German research consortium has released Soofi S 30B-A3B, an open-source, hybrid Mamba-Transformer Mixture-of-Experts foundation model. Trained on Deutsche Telekom's AI cloud using NVIDIA B200 GPUs, the model prioritizes German language data, significantly outperforming its architecture-identical reference on German benchmarks. While it shows competitive performance against other open models in English, larger models like Qwen3.5 35B-A3B still hold the top scores. The model's weights are available on Hugging Face, with custom code required for its use. AI
IMPACT This release contributes to the growing ecosystem of open-source models, particularly for non-English languages, potentially accelerating research and development in those areas.
RANK_REASON Release of a new open-source foundation model by a research consortium. [lever_c_demoted from research: ic=1 ai=1.0]
- Deutsche Telekom
- ellamind
- Fraunhofer Institute for Intelligent Analysis and Information Systems
- German Federal Ministry for Economic Affairs and Energy
- German Research Center for Artificial Intelligence
- Hugging Face
- KI Bundesverband
- Mamba-Transformer
- Nemotron 3 Nano
- NVIDIA
- Soofi Consortium
- Soofi S 30B-A3B
- Technical University of Darmstadt
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
