Kolibri has 78 billion total parameters, with approximately 3 billion active per token, according to Aleph Alpha. The German company released the model on 3 October 2026 for critical public administration and industry applications. Its description focuses on data control, compliance documentation and German-language operation.
How did Aleph Alpha develop Kolibri?
Aleph Alpha says Kolibri was developed as a sovereign AI model made in Germany for organisations that need to manage their data. The company says it trained the model to operate natively in German. German-language text accounts for around 23% of its pre-training data, according to Aleph Alpha.
The company was founded in 2019 and employs around 200 people across four locations in Germany, the source says. Aleph Alpha says its specialised LLMs are used by enterprises and government institutions seeking data sovereignty and trustworthy AI applications. The source does not identify Kolibri customers.
What does its model architecture mean?
Kolibri uses a mixture-of-experts (MoE) design, according to Aleph Alpha. Rather than activating every part of the model for each inference, the architecture selects relevant parts. The company says this supports efficiency without compromising capability.
The model has 78 billion total parameters, while approximately 3 billion are active per token. Parameters are components used by a model to process information; the active-per-token figure describes how many are engaged for each token, as stated in the source. Aleph Alpha does not provide independent performance measurements alongside these figures.
What can Kolibri do with documents?
Aleph Alpha says Kolibri supports retrieval-augmented generation (RAG), agentic workflows and native tool calling. In RAG, a model uses supplied documents to answer a question. The company says Kolibri is trained to decline an answer when those documents do not contain sufficient evidence.
Aleph Alpha says this is intended to reduce hallucinations in high-stakes settings. It also says built-in reasoning controls let users adjust the balance between reasoning depth and response speed. The source gives no test results showing how often the model abstains or how its answers perform.
What checks and documentation are described?
The company says it screened Kolibri against a blocklist of more than 4.5 million URLs. It also says each third-party dataset was reviewed for licence terms, lawful sourcing and adherence to opt-outs. Its technical report sets out how it addresses copyright, data protection and EU AI Act requirements.
Aleph Alpha says licensing terms and supporting material are available on Hugging Face. The source does not detail the terms or describe external review of the checks. CEO Ilhan Scheer said the company wants to make AI capability available in everyday operations in industry and public administration.
What remains to be established?
The supplied information does not name organisations that have deployed Kolibri or provide operational results. It also does not give a detailed account of the model’s licence or explain how customers will document compliance in practice. Aleph Alpha’s statements set out the intended use and technical approach, while evidence from deployments would show how those claims translate into use. The company says its ambition is to bring the capability into everyday operations.