Kolibri gives Europe an open-weights model that speaks German

Aleph Alpha has released an open-weights language model it calls sovereign, and it is built to speak German. Kolibri arrived on 3 October, the Day of German Reunification, with its full weights published on Hugging Face under the Apache 2.0 licence. The Heidelberg company's own tech blog describes it as an English-German Mixture-of-Experts Transformer with 78 billion total parameters, of which 3.46 billion are active per token.

A big model with a small working set

The architecture is what makes Kolibri affordable to run. Because it is a mixture-of-experts model, only a fraction of those 78 billion parameters fire for any given token, so the compute cost per answer sits closer to a small model than its headline size implies. The model card lists a context window of up to 1,048,576 tokens, though Aleph Alpha recommends staying at or below 262,144 for serving efficiency and complex tasks. Weights are stored in an 8-bit floating-point format, giving a memory footprint of roughly 78 GB, which the company says needs at minimum two A100 80 GB cards, or a single H200, B200 or B300.

Trained for regulated work, on 20 trillion tokens

Kolibri was pre-trained on 20 trillion tokens of a filtered, bilingual corpus that Aleph Alpha says is about 62.5 per cent English, 23.9 per cent German and 13.6 per cent code, followed by further mid-training and a long-context extension. Its knowledge cutoff is 18 June 2026. The model supports an explicit reasoning mode and tool calling, and the company positions it for mission-critical work in regulated sectors such as public administration, industry and aerospace.

Why sovereignty is the whole point

Aleph Alpha defines sovereignty on two fronts: how the model is built, and how it transfers to customers. The company says Kolibri comes with full supply-chain integrity, that every training decision is accounted for, and that customers get freedom of deployment and intellectual-property safety, so compliance is inherited rather than bolted on afterwards. Aleph Alpha is also a signatory of the EU's GPAI Code of Practice, which matters for public-sector buyers who have to justify a model choice on the record. Because the weights are open, the model can be run entirely on a customer's own hardware.

Our opinion

Kolibri is not trying to out-reason the frontier labs, and judging it that way misses the point. A model with 3.46 billion active parameters and German as a first-class language is aimed at a specific buyer: a government department, a hospital group or a manufacturer that cannot send its data to an American API and cannot afford to build its own model from scratch either. The Apache 2.0 licence is the sharp part of the offer, because it removes almost every legal excuse a procurement team could raise. Whether the model is genuinely good at German bureaucracy is a separate question that the open weights now let anyone test, and that accountability is exactly what the sovereignty pitch promises. Europe keeps saying it wants AI it controls; here is one it can actually download.