Aleph Alpha Launches Kolibri, a Sovereign AI Model for Government and Industry

Aleph Alpha Releases Kolibri

Aleph Alpha has launched Kolibri, a specialised large language model designed for mission-critical work in public administration, industry and aerospace.

Aleph Alpha positions Kolibri as a sovereign AI model, giving organisations greater control over how the technology is deployed and where sensitive information is processed. The company says the model was built in Germany and trained on infrastructure in Germany and Finland under European and German law.

78B Parameters With a Smaller Active Footprint

Kolibri uses a mixture-of-experts architecture with 78 billion total parameters, while only about 3.46 billion parameters are active per token. It supports context windows of up to 1 million tokens and is designed for both German and English.

The model's full weights are available through Hugging Face under the Apache 2.0 licence, allowing organisations to download and deploy the model on their own infrastructure. This is central to Aleph Alpha's sovereignty proposition, particularly for customers that do not want sensitive workloads sent to external AI inference services.

Designed for Government and Regulated Industries

Aleph Alpha has specialised Kolibri for reasoning, mathematics, coding, agentic tasks and long-context workloads, with additional attention to the requirements of regulated industries.

The company says the model is designed around use cases in public administration, industrial technology, automotive, semiconductors and aerospace. Its smaller active parameter count is intended to provide a balance between model capability and serving costs for organisations running AI on their own infrastructure.

German Language and Grounded Responses

A major part of Kolibri's positioning is its native German capability. Aleph Alpha says 21.3% of the model's pre-training tokens are German, with a bilingual German-English tokenizer and limited reliance on machine-translated material.

The company has also trained Kolibri to abstain when available information does not support an answer. Its Merlin-Arthur protocol and abstention training are intended to reduce unsupported responses in document-based applications, a feature Aleph Alpha sees as particularly relevant for government and enterprise workflows.

Sovereignty Extends Beyond the Model

Aleph Alpha's argument is that sovereign AI involves more than simply hosting a model inside Europe. The company says it controls the development pipeline from data curation and training through evaluation and deployment, while customers retain freedom over where and how Kolibri is operated.

The approach is aimed at organisations dealing with sensitive government, industrial or regulated data. Running the open-weight model on customer-controlled infrastructure can reduce dependence on external inference providers, although the actual level of sovereignty will still depend on an organisation's hardware, hosting, access controls and deployment practices.

Europe Pushes Its Own AI Infrastructure

Kolibri arrives as European governments and businesses increasingly look for AI systems that can be deployed under European legal and data-governance requirements. Aleph Alpha is competing in a market where open-weight models give organisations more options to operate AI locally rather than relying exclusively on large US or Chinese cloud platforms.

The launch does not mean Kolibri matches every frontier model across every benchmark. Aleph Alpha's published comparisons show competitive results against selected open-weight models, but those are company-reported evaluations and do not establish overall leadership. The more specific proposition is control: a German-English model that organisations can inspect, deploy and adapt on infrastructure they control.