Aleph Alpha Releases Kolibri Sovereign AI Model For Regulated Sectors
Aleph Alpha is positioning Kolibri around sovereignty and compliance rather than raw scale, and the company says it built and trained the entire pipeline without foreign control, which matters for buyers in regulated sectors who need to explain how a model reached its output.
Reporting from 1 source: GIGAZINE.
German AI startup Aleph Alpha announced the language model Kolibri on October 3, 2026, the Day of German Unity. The model has 78 billion total parameters and 3 billion active parameters per token, supports English and German, and has a context length of up to 1 million tokens. It is trained and built in Germany and Finland, and targets public administration, industry, and aerospace.
Kolibri keeps 3 billion active parameters of its 78 billion total active during each token, which Aleph Alpha frames as a balance between model capability and deployment cost. The company says the model performs comparably to models with up to four times the active parameters, such as Nemotron 3 Super, on mathematics, coding, grounding, and long-context tasks.
On construction, Aleph Alpha points to the Merlin-Arthur protocol, an in-house method it says makes Kolibri abstain when context does not support an answer. Training ran on infrastructure in Germany and Finland under European and German law, and the model ships with control transferred to the customer for deployment and cost-latency tradeoffs.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.