German AI startup Aleph Alpha has launched Kolibri, a 78-billion-parameter open-weight language model optimized for German and English, according to the company. The model utilizes a mixture-of-experts architecture with approximately three billion active parameters per token. Kolibri was trained on 768 Nvidia B200 GPUs in Germany and Finland, featuring a 1-million-token context window and an Apache 2.0 open-source license.
Aleph Alpha claims Kolibri sits on the Pareto front of quality and operating cost in both supported languages, outperforming comparable models. German accounts for 21.3% of its training data via a dedicated pipeline, while synthetic data from Chinese models was also incorporated. The model targets applications in public administration, aviation, and industrial sectors under compliance with the European Union AI Act.
The release underscores growing efforts toward European digital sovereignty in artificial intelligence. By open-sourcing model weights on Hugging Face, Aleph Alpha aims to provide regional enterprise and government entities with locally hosted, compliant AI infrastructure that reduces reliance on foreign proprietary providers.
Why it matters
Enterprise adopters gain an open-weight 1M-context model built specifically for EU AI Act compliance.
Efficient MoE architecture lowers operational costs by activating only three billion parameters per token.
Regional focus on specialized non-English data pipelines provides competitive advantages in localized public sector deployments.
Source: the-decoder.com



