Aleph Alpha releases Kolibri-1, a 78-billion-parameter open-weight model for English and German, on German Reunification Day

German AI company Aleph Alpha released Kolibri-1 on 3 October. Its blog post opens: “On the Day of German Reunification, we are releasing our new model: Kolibri.” It calls Kolibri “a specialized language model built for sovereign mission-critical work in regulated areas including public administration, industrials and aerospace.”

The model card on Hugging Face gives the details. Kolibri-1 has about 78 billion parameters in total, of which about 3.46 billion are active for each token: it is a mixture-of-experts model with 384 routed experts per layer, six chosen per token. It accepts up to 1,048,576 tokens of context. It was pre-trained on 20 trillion tokens on 768 Nvidia B200 GPUs, and is released under the Apache 2.0 licence, so anyone can download and run it.

Aleph Alpha's own scores, as stated on the card, include 84.3 on GPQA Diamond in English and 81.3 in German, and 85.9 on LiveCodeBench v6. These are the company's figures, not an outside test. The card describes training the model to abstain when it does not have enough information, and says Aleph Alpha has signed the EU's code of practice for general-purpose AI.

A model this sparse is cheap to serve for its size, which matters to public bodies and companies that want to run AI on their own hardware rather than send data to a US or Chinese provider. As of Sunday afternoon none of three model-release trackers checked had listed it.

• Confirmed Kolibri-1 has about 78B total and 3.46B active parameters, a 1,048,576-token context and an Apache 2.0 licence. Aleph Alpha model card, Hugging Face
• Claimed Aleph Alpha reports 84.3 (English) and 81.3 (German) on GPQA Diamond and 85.9 on LiveCodeBench v6. Aleph Alpha model card, Hugging Face
• Confirmed Aleph Alpha released Kolibri on 3 October, German Reunification Day, for regulated uses including public administration, industrials and aerospace. Aleph Alpha