Aleph Alpha Kolibri: Open-Weight German-English Model Released Under Apache 2.0

What happened

Aleph Alpha has released Kolibri, an open-weight, German-English language model built on a mixture-of-experts architecture with 78 billion parameters, of which around three billion are active for each token processed. The company made the weights freely available under the Apache 2.0 licence, which allows broad commercial and research use.

According to The Decoder's reporting, more than 21 percent of Kolibri's training data is German, a deliberate weighting for a model aimed at German-speaking users and organisations. Training ran on 768 B200 GPUs split across facilities in Germany and Finland, a detail that places the model's development inside Europe's own compute infrastructure.

The release is framed by its developers and by The Decoder as a case for European AI sovereignty: the ability of organisations in the region to build on capable models without depending entirely on overseas platforms.

Why it matters

For leaders in digital transformation, the significance lies in choice. A mixture-of-experts design that activates only a fraction of its parameters per token is built to balance capability against the cost of running it. An open-weight release under a permissive licence means organisations can inspect, fine-tune and deploy the model on infrastructure they control, which changes the conversation from "which vendor do we rent?" to "what do we need the model to do, and where should it run?"

Language weighting matters too. A model trained with a substantial share of German text is better placed to handle the idiom, regulatory vocabulary and service conversations that German-speaking customers and employees actually use. For public-sector bodies, banks, insurers and telecoms operating in German, that makes a locally tuned foundation model a credible option to evaluate alongside larger general-purpose systems.

By the numbers

• 78 billion parameters in total, with roughly 3 billion active per token.
• More than 21 percent of the training data is German.
• 768 B200 GPUs were used for training, across Germany and Finland.
• Apache 2.0 licence, with weights freely available.

The Renascence take

Most coverage will read Kolibri as a sovereignty statement. The more useful reading is operational: the real sovereignty question is who controls the experience layer, not just the weights. A model you can run locally still produces the answers your customers hear, and those answers are shaped by the data, prompts, guardrails and escalation paths you design around it.

Owning the model does not give you a better customer experience; it gives you the right to fail in private and fix in public. The behavioural principle at play is trust, and trust is built on consistency, not provenance. Operators should treat an open-weight release as a reason to map their journeys first, pick the moments where language fluency genuinely changes outcomes, and pilot Kolibri-class models there with clear human hand-off rules. If the experience cannot be measured, no licence or data-residency claim will rescue it.