Aleph Alpha Launches Kolibri Model with 78 Billion Parameters

On October 5, 2026, Aleph Alpha released the Kolibri model, which boasts 78 billion parameters. This German-English language model utilizes approximately 3 billion active parameters per token and was trained on a dataset where 21.3% is German. Kolibri achieved a score of 71% on German benchmarks, indicating its effectiveness in that language. The model was trained using 768 B200 GPUs, allowing it to support context windows of up to one million tokens.

Aleph Alpha claims that Kolibri targets sectors such as public administration, aviation, and industry, positioning it as a versatile tool for various applications. The model weights are available on Hugging Face under an Apache 2.0 license, facilitating accessibility for developers and researchers. This release aligns with the ongoing discussions around the EU AI Act, which influences AI model development in Europe. This follows previous announcements from Aleph Alpha regarding their commitment to European AI sovereignty.