A German research consortium released Soofi S 30B-A3B, an open language model trained entirely on Deutsche Telekom's Munich cloud infrastructure. The model contains 31.6 billion parameters but uses a hybrid architecture that activates only a fraction of them per token, maintaining consistent throughput across long contexts.
The consortium deliberately weighted the training dataset toward German, and Soofi S now ranks at the top of benchmarks for both English and German language tasks. This positions it as a competitive alternative to larger closed models while requiring fewer computational resources during inference. The efficient parameter activation approach addresses a core constraint in deploying large language models: the ability to run them without prohibitive hardware costs.
The model's design reflects a broader European strategy to build AI capabilities independent of U.S.-controlled systems. By training entirely within Deutsche Telekom's infrastructure, the consortium ensured data processing remained within German jurisdiction, addressing regulatory and sovereignty concerns that increasingly shape European AI development.
Soofi S performs well on standard benchmarks covering reasoning, knowledge, and language understanding tasks across both languages. The bilingual focus reflects a practical market need. Many European AI applications require strong performance in local languages, yet most open models concentrate training data on English, leaving German and other European languages underrepresented.
The open release matters for European researchers and businesses building applications that need German language support. Companies can now fine-tune or deploy Soofi S without depending on commercial providers like OpenAI or Anthropic, reducing latency and cost while keeping data within European legal frameworks.
The efficient architecture also has implications beyond German applications. The technique of activating only necessary parameters scales to other language pairs and domains, potentially influencing how future models balance capability with practical deployment constraints. For teams building European AI infrastructure, Soofi S demonstrates that competitive open models can emerge from regional research efforts backed by established telecom infrastructure.