Aleph Alpha Releases Kolibri, an Open-Weight German-English AI Model Built for Sovereign Deployment

Aleph Alpha has released Kolibri, a new open-weight language model the Heidelberg-based company says is built specifically for regulated, mission-critical work in sectors such as public administration, industrials and aerospace. The release, timed to coincide with the Day of German Reunification, is being made freely available on Hugging Face under an Apache 2.0 license.

Kolibri is a Mixture-of-Experts transformer with 78 billion total parameters, of which only about 3 billion are active for any given task. That design keeps running costs down while supporting context windows of up to one million tokens. The model supports both English and German natively, with a tokenizer and training data built around the two languages rather than treating German as an afterthought.

The release follows an earlier, smaller model called Kolibri Origin, a 30-billion-parameter system Aleph Alpha used to validate its training pipeline before scaling up. The gap between the two releases was just three months, during which the company says it nearly tripled the parameter count, expanded the context window from 65,000 to a million tokens, and increased training data from 7.5 trillion to 20 trillion tokens. Aleph Alpha attributes the pace to a heavily automated "Model Factory" pipeline that runs training, evaluation and recovery from hardware failures without human intervention, logging checkpoints roughly every hour.

Aleph Alpha is pitching Kolibri less on raw scale and more on efficiency, specialization and sovereignty. The company says the model sits on the "Pareto frontier" of quality versus serving cost in both English and German, matching the performance of models up to four times its active parameter size, such as Nvidia's Nemotron 3 Super, on tasks spanning math, coding, long-context reasoning and agentic tool use.

A central claim is around German-language capability. Aleph Alpha says it avoided leaning on machine translation, which it argues imports the "cultural fingerprint" of English source text, and instead built roughly 2.4 trillion unique German tokens through a custom web-crawling pipeline and an internal rephrasing technique that rewrites existing German text into new forms, such as encyclopedia entries or dialogues, without translating from another language. German made up just over a fifth of the pre-training data, the company says.

The company also emphasises what it calls "grounding": training the model to say "I don't know" rather than guess when an answer isn't supported by the available context. It developed an in-house method called the Merlin-Arthur protocol, a training setup involving three roles that teaches the model to distinguish between contexts that support an answer and those that have been stripped of supporting evidence. Aleph Alpha reports that Kolibri abstains instead of answering wrongly far more often than its predecessor on public hallucination benchmarks.

Sovereignty is the other pillar of the pitch. Aleph Alpha says Kolibri was built and trained entirely in Germany and Finland under European law, with full visibility into its training data and process, and that customers get full freedom to deploy the model on their own infrastructure rather than send data to third-party cloud services. The company frames this as central to compliance with the EU AI Act, the EU's General-Purpose AI Code of Practice and GDPR.

Beyond public benchmarks, Aleph Alpha says it built internal evaluation suites tailored to specific industries, including the German public sector, aerospace, semiconductors, automotive supply and industrial drive technology, using synthetic training data modeled on customer workflows without training directly on customer data. The company reports performance gains across all five sectors through successive training runs.

Kolibri is available now for download, with instructions for deployment via the company's inference tooling. Aleph Alpha says it is directing larger organizations with specialization or deployment needs to its enterprise sales team.