Aleph Alpha, based in Heidelberg, has launched Kolibri, an open-weight AI model featuring 78.1 billion parameters. The model utilizes a Mixture-of-Experts architecture, where approximately 3.46 billion parameters remain active per token. It is specifically designed for regulated industries, including defense, aerospace, public administration, and general industrial sectors.
A key technical capability of Kolibri is its context window, which supports up to 1,048,576 tokens. The model was trained on a dataset ranging from 20 to 24 trillion tokens, with a significant focus on German-language content. It is optimized for performance in both English and German languages.
The training process took place on infrastructure located in Germany and Finland, utilizing 768 NVIDIA B200 GPUs. Aleph Alpha has also released a 189-page technical report detailing the model's architecture and training methodology. According to the company, the model achieved a score of 96.9% on the AIME 2025 math benchmark.
