On Germany’s Day of Reunification, Aleph Alpha released Kolibri — an English-German Mixture-of-Experts Transformer with 78B total parameters and 3B active, supporting context lengths up to 1M tokens. The model is available on Hugging Face under Apache 2.0 license terms.
Kolibri follows Kolibri Origin (30B total, 3B active, 65k context), validating Aleph Alpha’s model training pipeline. The company reports hundreds of ablation experiments, stable pre-training resilient to hardware failures, and continuous monitoring of training metrics and custom benchmarks.
Kolibri is specialized for sovereign mission-critical work in regulated areas including public administration, industrials, and aerospace, optimized for German language, reasoning, math, and agentic behavior. Aleph Alpha emphasizes sovereignty: full supply-chain integrity, transparency, freedom of deployment, and IP safety for customers.
Benchmark claims: Kolibri sits on the Pareto frontier for quality versus serving cost in English and German. It matches models with up to four times its active parameter count on math, coding, grounding, and long-context tasks. Examples include AIME 2025 at 96.9, GPQA diamond at 84.3, HumanEval+ at 92.7.
The model was built with EU AI Act, GPAI Code of Practice, and GDPR in mind. Training occurred in Germany and Finland under European law. 21.3% of pre-training tokens are German with a bilingual tokenizer designed for organic German data rather than translation-heavy corpora.
Pipeline timeline: Kolibri Origin finished pre-training June 11, 2026; Kolibri finished September 11, 2026. Between releases the team scaled from 30B to 78B parameters, 65k to 1M context, 7.5T to 20T training tokens, processing over 200T raw tokens through curation.