Aleph Alpha released the open AI language model Kolibri on October 3, 2026. The company from Heidelberg trained the model with 78.1 billion parameters entirely in Germany and Finland and provides the weights for free under the Apache 2.0 license.
The target audience includes government agencies, industry, and aviation companies that want to operate AI on their own infrastructure.
Model processes up to one million tokens of context
Kolibri is a mixture-of-experts model with 384 experts, of which six are active per request. This means the system effectively only computes with 3.46 billion parameters per token, although a total of 78.1 billion parameters are stored.
A proprietary tokenizer called UniBPE is designed for both German and English, and four levels of reasoning effort can be set, ranging from no additional reasoning to intensive multi-step reasoning.
Kolibri was trained with approximately 24 trillion tokens, of which 62 percent are in English, about 21 percent in German, and 14 percent in code. According to the company, the context window extends up to one million tokens, though it was actually trained up to 256,000 tokens.
For pre-training, 768 Nvidia B200 graphics cards ran for 21 days, completed on September 11, 2026. An additional training method based on the so-called Merlin-Arthur protocol is meant to reduce the model's tendency to fabricate answers.
The predecessor was Kolibri Origin, with 30.6 billion parameters, released just three months earlier via the company's own "Model Factory" infrastructure for rapid model iterations.
Aleph Alpha positions Kolibri as a response to US cloud dependency
Co-founder Samuel Weinbach says, according to WirtschaftsWoche, that the model can specifically "argue in German," making it practically usable for everyday work in government agencies.
Co-CEO Ilhan Scheer describes AI sovereignty as the ability to make decisions over one's own technology instead of staying permanently dependent on US cloud providers. Digital Minister Karsten Wildberger (CDU), who had already publicly backed the Aleph Alpha collaboration before, also attended the presentation in Heidelberg.
Target groups are primarily government, industry, and aviation – sectors with high demands on data protection and reliability that are reluctant to rely on US cloud services. The merger with Cohere sealed in September was meant to turn Heidelberg into a pure research site, while operational business moves to Toronto and Berlin.
With Kolibri, the roughly 200-person Heidelberg team now shows off a market-ready model it trained entirely itself – more than basic research. The Cohere merger still faces antitrust review in several jurisdictions, and Heidelberg remains operationally independent until it closes.
Independent tests show mid-pack performance and heavy hardware needs
An independent analysis by Orcarouter ranks Kolibri behind the open rival model Qwen3.8, which has 27 billion parameters, on general language understanding, where Qwen3.8 reaches higher average scores in both English and German.
On a German-language math test (AIME 2025), however, Kolibri scores 87.5, beating several larger models in the same comparison group.
The same analysis lists a minimum of two Nvidia H100 graphics cards, or comparable hardware with roughly 78 gigabytes of combined memory, as the baseline requirement – about the same size as the compressed model itself. Aleph Alpha does not offer hosted access through its own API at launch; interested users download the weights from Hugging Face and run Kolibri on their own or rented servers, a hurdle that is considerably higher for smaller agencies without their own data centers.
It remains open whether Aleph Alpha will also offer Kolibri as a hosted service through the sovereign cloud it built together with the Schwarz Group. For agencies without their own GPU clusters, that would be the real test of whether the open model generates genuine demand beyond pilot projects.
