An open-source language model called Kolibri has been released on the Day of German Reunification, highlighting a push toward sovereign artificial intelligence in regulated sectors. Kolibri is an English-German Mixture-of-Experts Transformer with 78 billion total parameters and 3 billion active parameters, and it supports context lengths of up to 1 million tokens. The model is available for download with full weights on Hugging Face under the Apache 2.0 license.
Purpose and Target Sectors
Kolibri is built for sovereign mission-critical work in regulated areas including public administration, industry, and aerospace. It has been specialized for German, reasoning, math, agentic behavior, and further capabilities that customers need in production. The specialized design aims to improve contextualized performance and help customers monitor measurable return on investment over time.
Training Pipeline and Model Iteration
The release follows a continuous iteration of the organization's model training effort, which was first validated with Kolibri Origin. Kolibri Origin featured 30 billion total parameters, 3 billion active parameters, and a shorter 65,000 token context window. The same pipeline enabled hundreds of ablation experiments and stable pre-training that continued through hardware failures or data connection drops without manual intervention.
Sovereignty and Supply-Chain Integrity
Sovereignty is framed as a combination of how the model was built and how it transfers to customers. The team provides full supply-chain integrity, from data ingestion through pre-training, post-training, and final evaluations. Customers retain full freedom of deployment and intellectual-property safety, with compliance positioned as an inherited property of the model.
Efficiency and Performance Trade-Offs
Kolibri optimizes the trade-off between model capability and deployment costs by using 3 billion active parameters out of 78 billion total. It is described as sitting on the Pareto frontier for quality versus serving cost in both English and German. Across math, coding, grounding, and long-context tasks, the model matches systems with up to four times its active parameter count, such as Nemotron 3 Super.
Contextualized Industry Evaluation
Because public benchmarks often fail to capture specialized sector needs, the team developed internal evaluation suites for the German public sector, aviation, manufacturing, and automotive industries. Each suite mirrors the skills, workflows, and edge cases required in these verticals, and paired synthetic training environments helped improve Kolibri without using customer data. The model was also trained with abstention data and a Merlin-Arthur protocol to say I don't know when context does not support an answer.
Bilingual Design and German Data
A bilingual German and English tokenizer supports Kolibri, with German accounting for 21.3% of pre-training tokens. Translation was used sparingly at 6% overall because translated text tends to carry the cultural fingerprint of its source language. This approach produces a model that is bilingual by design rather than an English model that has read some German.
Compliance and European Control
The model was developed with the EU AI Act, the General-Purpose AI Code of Practice, and the GDPR in mind from the ground up. Developers highlight transparency around model weights and training data curation, while reasoning traces make it possible to understand how the model reached a particular answer. Kolibri was built in Germany and trained on infrastructure in Germany and Finland under European and German law, with no foreign control.
Kolibri's release underscores the increasing importance of efficient, transparent, and sovereign language models for regulated industries. By combining a small active parameter footprint with long-context support and specialized bilingual capabilities, the model targets on-premises deployment without relying on third-party inference services. Its open availability under the Apache 2.0 license may encourage organizations to evaluate it for compliance-sensitive workflows and mission-critical AI operations.