Aleph Alpha releases Kolibri, an open-weight German-English model

Aleph Alpha made Kolibri, a specialised large language model for public-administration and industrial use, available from 3 October 2026 — Germany’s national Unity Day — with the company’s own announcement following two days later. The model combines 78 billion total parameters with around 3 billion active per token in a mixture-of-experts design, supports German and English, and uses a tokenizer the company said was tailored to represent German text more efficiently than general-purpose tokenizers.

Aleph Alpha said German-language text made up around 23% of Kolibri’s pre-training data, and that the model supports agentic workflows, retrieval-augmented generation and native tool calling, including being trained to abstain from answering when supplied documents lack sufficient evidence. The company published a technical report alongside the model documenting compliance measures aimed at the EU AI Act and copyright law, saying all training data was screened against a blocklist of more than 4.5 million URLs, including sources drawn from the European Commission’s Piracy Watch List, and that each third-party dataset was checked for licensing terms and lawful sourcing before use. Chief executive Ilhan Scheer framed the release as proof that “we have the talent and expertise in Germany to develop competitive AI models.” No independent benchmark results or training-token totals were disclosed.

Weights and “applicable license terms” were made available on Hugging Face, with Aleph Alpha not specifying a standard open-source licence by name in its announcement.

The release came weeks after Aleph Alpha and Canada’s Cohere signed an agreement to merge into a “transatlantic sovereign AI” company; Aleph Alpha said in the Kolibri announcement that it continues to operate independently until that deal closes, pending regulatory approval. Kolibri is therefore likely the last model Aleph Alpha ships under its own name before any merger completes, and serves as a demonstration of the sovereignty pitch — government and regulated-industry customers controlling their own deployment infrastructure — that underpins the Cohere combination.