Aleph Alpha launches Kolibri, a new large language model for German and English

Kolibri is a large language model (LLM) with 78 billion parameters, utilizing a mixture of experts architecture that activates approximately 3.5 billion parameters per token processed.

The model was released on October 3, 2026, under the Apache 2.0 license and is available on Hugging Face. It was trained in Germany and Finland, adhering to European laws, and aims to comply with the EU AI Act.

Aleph Alpha describes Kolibri as 'sovereign,' meaning it can be deployed on local infrastructure without external control, ensuring data privacy and intellectual property safety.

Kolibri's tokenizer is optimized for German, requiring 11.2% fewer tokens than competitors for German text, which enhances efficiency in processing.

The model supports a context length of up to 1 million tokens and employs a unique reasoning method that allows it to acknowledge when it does not know an answer, reducing the risk of hallucination.

Kolibri is particularly suited for applications requiring German language processing and data privacy, making it ideal for public authorities and organizations in sectors like healthcare and manufacturing.

However, it requires significant GPU resources, specifically around 78 GB of memory, limiting its accessibility for smaller operations.