Aleph Alpha released Kolibri, an open-weight language model designed for sovereign, mission-critical applications in regulated industries. The model is a Mixture-of-Experts Transformer with 78 billion total parameters and 3 billion active parameters, supporting context lengths up to 1 million tokens. It is available for download with full weights on Hugging Face under Apache 2.0 license terms.
Kolibri was built as a bilingual English-German model, with 21.3% of pre-training tokens in German and only 6% translated text. According to Aleph Alpha, this approach creates a model that is “bilingual by design, not an English model that has read some German.”
The model is optimized for performance in specific sectors including public administration, industrial operations, and aerospace. Aleph Alpha specialized Kolibri for German, reasoning, math, and agentic behavior tailored to customer needs. The company positions Kolibri on what it calls the “Pareto frontier for quality versus serving cost,” claiming it matches performance of models with up to four times its active parameter count on tasks including math, coding, grounding, and long-context work.
According to Aleph Alpha’s benchmarks, Kolibri scores 96.9 on AIME 2025 and 84.3 on GPQA (diamond). On domain-specific internal benchmarks developed for aviation and retail sectors, Kolibri scored 76.7 on an airline benchmark and 69.9 on a retail benchmark.
Sovereignty is central to Kolibri’s design. Aleph Alpha built the model in Germany, trained it on infrastructure in Germany and Finland under European law, and owns the entire pipeline from data curation through training and optimization. The company developed the model with the EU AI Act and GDPR in mind. Customers can deploy Kolibri on-premise without sending internal data to third-party services.
Kolibri features grounding capabilities, trained to abstain from answering when information is not present in provided context. The model uses Aleph Alpha’s Merlin-Arthur protocol to ensure it refrains from answering unsupported questions.
Development followed Aleph Alpha’s Model Factory approach. The company first released Kolibri Origin, a 30-billion-parameter model with a 65,000-token context window, in June. Kolibri was completed three months later in September, demonstrating what Aleph Alpha describes as rapid iteration through a standardized training pipeline.
Key facts
- Kolibri has 78B total parameters with 3B active, supporting up to 1M token context lengths
- The model is bilingual, with 21.3% German tokens in pre-training
- Available under Apache 2.0 license on Hugging Face
- Built and trained entirely in Germany and Finland under European law
- Scores 96.9 on AIME 2025 math benchmark and 84.3 on GPQA diamond benchmark
- Designed for regulated sectors including public administration, aerospace, and manufacturing
- Trained to abstain from answering questions not supported by provided context