Kolibri Has Landed: A Sovereign Open-Weight Model

Aleph Alpha released Kolibri, a bilingual English-German Mixture-of-Experts Transformer with 78 billion total parameters and 3 billion active parameters, supporting context windows up to 1 million tokens. The full weights are available on Hugging Face under Apache 2.0. Kolibri is positioned for sovereign, regulated use in public administration, industrials and aerospace, with specialization for German, reasoning, math, coding and agentic tasks. Development used a Model Factory pipeline that enabled rapid iteration from Kolibri Origin (30B total, 3B active, 65k context) to Kolibri in three months: increase to 78B, extended context, and growth from 7.5T to 20T training tokens distilled from over 200T raw tokens. Architectural changes included new attention design, tripled experts, greater sparsity, revised routing, expanded post-training data and multi-level reasoning training.

Benchmarks show Kolibri matches or outperforms models with up to four times its active parameters on math, code, grounding and long-context tasks, and is claimed to sit on the Pareto frontier of quality versus serving cost. Internal customer-proxy evaluations report large gains across verticals (e.g., automotive supplier 0.72→0.99, aerospace 0.14→0.59). Training emphasized grounding and abstention via a Merlin-Arthur protocol so the model can refuse unsupported answers. About 21.3% of pretraining tokens are German (6% translation), and development and training occurred under European control to ensure supply-chain transparency, explainability, deployment freedom and compliance with EU regulations.