The efficiency claim rests on a mixture-of-experts design: Kolibri has 78 billion parameters in total, but only about three billion are active for any given token, which is how the company says it keeps quality up while decoding faster than comparable models. Aleph Alpha claims the model sits on the Pareto front of quality and operating cost in both languages, meaning it says no similarly architected rival beats it on one of those measures without losing ground on the other. That comparison includes models released in March and April 2026, some of which it says it outperforms significantly.

German accounts for 21.3 percent of the training data, supported by a dedicated German data pipeline built for the project, and Chinese models were used to generate synthetic training data. The company reports a 71 percent score on German benchmarks.

Kolibri targets public administration, aviation and industry, and was developed under European law with the EU AI Act in mind. It supports context windows of up to one million tokens and was trained on 768 Nvidia B200 GPUs in Germany and Finland, according to the tech report. The weights are available under an Apache 2.0 license on Hugging Face.