AI from Germany: Aleph Alpha introduces open-weight model Kolibri

Aleph Alpha has released a new language model named Kolibri. The Mixture-of-Experts model has a total of 78.1 billion parameters, of which around 3.46 billion are active during the processing of a token. According to Aleph Alpha, this architecture is intended to enable performance comparable to larger models with lower computational effort. The comparison values come from Aleph Alpha itself; independent confirmation of the performance data is still pending.

Kolibri supports German and English, logical reasoning, and the use of external programs and services, as well as context lengths of up to around one million tokens. According to Aleph Alpha, Kolibri was specifically tailored for German-language applications: around 23 percent of the pre-training tokens are German. Furthermore, Kolibri has been optimized to efficiently process German texts and to reason in German.

Kolibri is an open-weight model. The model weights and configuration files are available for download on Hugging Face under the Apache 2.0 license. According to the Model Card, the license does not apply to the underlying code, model architecture, and training methods.

Kolibri aims to reduce dependencies

Aleph Alpha positions Kolibri as a model for business-critical applications in companies, authorities, and other regulated sectors. It is suitable for processing long documents, programming tasks, and agentic workflows, among other things.

The startup explicitly markets Kolibri as a “sovereign” AI model. By this, Aleph Alpha means control over the development chain from data preparation to training on the one hand, and control at the customer’s end on the other: Kolibri can be operated on the customer’s own infrastructure, so that internal data does not have to be transferred to external services. According to Aleph Alpha, the model was developed in Germany and trained on infrastructure in Germany and Finland. However, Aleph Alpha also used external models to generate synthetic training data. The technical report mentions GLM-5.2 and GLM-5.3 from Z.ai, as well as Qwen3.8-27B.

Aleph Alpha also refers to measures regarding copyright, data protection, and the EU AI Act. According to the company, all training data was checked against a blocklist of more than 4.5 million URLs, which is based, among other things, on the European Commission’s Piracy Watch List.

Aleph Alpha faces merger with Cohere

Aleph Alpha was founded in 2019 and is headquartered in Heidelberg. The company employs around 200 people, according to its own statements. In 2023, Aleph Alpha announced a financing package of 470 million Euros. The Schwarz Group and Bosch Ventures, among others, participated in the financing round.

In September 2026, Aleph Alpha agreed to a merger with the Canadian AI company Cohere. The joint company will operate worldwide under the name Cohere. The transaction is still subject to regulatory approvals.

(dahe)