Aleph Alpha releases Kolibri, an open-weight German-English model aimed at European public administration and industry

Aleph Alpha has released Kolibri, a German-English language model with 78 billion parameters, of which about 3 billion are active per token through a mixture-of-experts architecture. The weights are available on Hugging Face under an Apache 2.0 license. The company says Kolibri sits on the Pareto front of quality and operating cost in both languages, and that it scores 71 percent on German benchmarks while decoding faster than comparable models such as GPT OSS A5B, Qwen 3.6 A3B and Gemma 4 A4B. The Decoder notes that some of those comparison models date from March and April 2026. German makes up 21.3 percent of the training data, supported by a dedicated German data pipeline, and Chinese models were used to generate synthetic training data. According to the technical report, Kolibri supports context windows of up to one million tokens and was trained on 768 B200 GPUs in Germany and Finland. Aleph Alpha says it developed the model under European law with the EU AI Act in mind. It is targeting public administration, aviation and industry, where buyers may prefer a model they can run under their own control.