Aleph Alpha

AI/LLM · Sovereign AI · Enterprise Software

Launch

Aleph Alpha releases Kolibri, a sovereign MoE model with 1M-token context

It arrives weeks after the company agreed to a Cohere tie-up widely seen as an acquisition, testing its sovereignty claims.

• Aleph Alpha released Kolibri on October 3, 2026, an open-weight Mixture-of-Experts model with 78B total and roughly 3B active parameters, licensed under Apache 2.0 with full weights on Hugging Face.
• The Heidelberg-based company built Kolibri for sovereign, mission-critical use in regulated sectors like public administration, industrials and aerospace, and employs about 200 people across four German sites.
• Kolibri natively handles 262,144 tokens of context and was tested up to 1,048,576 tokens, trained on about 24 trillion tokens (over a fifth German) using 768 Nvidia B200 GPUs.
• The model was trained from scratch on infrastructure located in Germany and Finland, with Aleph Alpha keeping its training code and methods while open-sourcing only the weights and configuration files.
• Co-founder Samuel Weinbach said Kolibri was tailored to German-language requirements so it can reason fluently in German, which executives frame as key to everyday enterprise workflows.
• The launch follows an April agreement between Aleph Alpha and Canadian AI firm Cohere to build a transatlantic sovereign-AI venture, a deal widely described as a de facto acquisition of Aleph Alpha.
• Kolibri pushes back on the idea that heavy EU regulation stifles innovation, betting that built-in compliance and customer-controlled infrastructure can win government and industrial buyers wary of US cloud dependence.