Aleph Alpha Releases Kolibri, an Open-Weight 78B MoE Model Aimed at European Sovereignty

German AI developer Aleph Alpha has officially released Kolibri, an open-weight bilingual language model designed to champion European technological sovereignty. Released under the permissive Apache 2.0 license on Hugging Face, the model targets enterprise and public sector use cases across Europe.

Kolibri features a mixture-of-experts (MoE) architecture with 78 billion total parameters, activating approximately 3 billion parameters per token during inference. The architecture is engineered to balance performance and inference cost, with Aleph Alpha reporting that the model sits on the Pareto front of quality and operational expense for both German and English.

Benchmark Performance and Architecture

Kolibri demonstrates competitive speed and accuracy, scoring 71% on German language benchmarks. According to Aleph Alpha, the model achieves faster decoding speeds than comparable lightweight active-parameter architectures, including GPT OSS A5B, Qwen 3.6 A3B, and Gemma 4 A4B.

In addition to its MoE structure, Kolibri supports an extensive context window of up to one million tokens, enabling the processing and retrieval of vast documents and enterprise data repositories in a single prompt.

Training Infrastructure and Data Pipeline

To build Kolibri, Aleph Alpha trained the model across 768 Nvidia B200 GPUs hosted in data centers located in Germany and Finland.

German-language content accounts for 21.3 percent of the pre-training dataset, supported by a specialized data pipeline developed specifically for this project. The company also utilized Chinese language models to generate synthetic training data to enhance the model's capabilities.

Enterprise and Regulatory Focus

Kolibri is explicitly tailored for sensitive and highly regulated sectors, primarily targeting public administration, aviation, and industrial applications.

A central selling point for Aleph Alpha is compliance: Kolibri was developed under European legal frameworks and built from the ground up with the requirements of the EU AI Act in mind. By providing an open-weight architecture paired with strict regulatory alignment, the release offers European organizations an alternative to proprietary models hosted outside the region.