Aleph Alpha released Kolibri, a sovereign large language model (LLM) with 78 billion parameters, on October 3, 2026. The model supports German and English and operates under the Apache 2.0 license. Kolibri’s weights are publicly available on Hugging Face, and it was trained from scratch using infrastructure located in Germany and Finland, reflecting a focus on data privacy and regulatory compliance.
Kolibri uses a mixture of experts architecture, activating only about 3.5 billion parameters per token generated or processed. This design enables efficient computation while maintaining high performance. The model was developed with the EU AI Act in mind, aiming to balance innovation with regulatory requirements on data privacy, stewardship, and environmental impact. Aleph Alpha’s evaluation indicates Kolibri outperforms comparable models of similar size in both German and English, according to a detailed 189-page technical report and model card.
The launch of Kolibri underscores Germany’s effort to build sovereign AI technology that complies with strict European regulations. It contrasts with perceptions that regulation stifles innovation, as Kolibri demonstrates a model built from the ground up to meet these standards. The availability of Kolibri’s open weights supports transparency and wider research use, positioning it alongside other large-scale models while emphasizing ethical and environmental considerations.
Kolibri’s release on October 3, 2026, marks a significant milestone for European AI development. The model is accessible on Hugging Face, allowing developers and researchers to deploy and experiment with it under the Apache 2.0 license, fostering further innovation within regulatory frameworks.