Aleph Alpha’s Kolibri: Germany’s Open-Weight “Sovereign” AI Model Is Ready to Download

October 4, 2026 — Heidelberg, Germany

German AI company Aleph Alpha released Kolibri on October 3, a new open-weight language model built for sovereign, on-premise use in government and regulated industries — and anyone can download the full weights on Hugging Face under the Apache 2.0 license. Why it matters: it’s a rare serious open model built end to end in Europe, optimized for German and English, and designed to run on an organization’s own servers rather than a US or Chinese cloud.

What happened

Aleph Alpha launched Kolibri on the Day of German Reunification, a deliberate bit of symbolism. The company calls it a “sovereign” model, and means two things by that: it was built by teams in Germany, trained on infrastructure in Germany and Finland under European and German law, with no foreign control — and customers get full freedom of deployment, so compliance comes as an inherited property of where and how the model runs.

The details

• Architecture: an English-German Mixture-of-Experts transformer with 78.1 billion total parameters, but only about 3.46 billion active per token — enough capacity for serious work while staying cheap to serve.
• Context: trained natively to 262,144 tokens, with settings to serve up to 1,048,576 tokens of context.
• License: Apache 2.0 for the weights and configuration files — commercial use, modification, and redistribution permitted. BF16 and FP8 checkpoints are on Hugging Face.
• Capabilities: four reasoning effort levels (none, low, medium, high), tool calling, and specialization for German, math, coding, long-context work, and agentic tasks.
• Safety design: Aleph Alpha built the model with the EU AI Act, the EU’s General-Purpose AI Code of Practice, and GDPR in mind. It uses a “Merlin-Arthur” approach that trains the model to refuse to answer when the provided context is insufficient to verify information — a deliberate anti-hallucination stance.

Kolibri is the successor to Kolibri Origin, a smaller 30B-total model the company used to validate its automated training pipeline. The full model was trained on roughly 24 trillion tokens, about 21.3% of them German, and Aleph Alpha says it runs on a single GPU-class card (or two H100s) through standard serving stacks.

Why it matters

Most frontier models worth deploying live inside American or Chinese infrastructure, and European governments are increasingly uncomfortable with that. Kolibri is positioned squarely at that anxiety: a ministry or car supplier can run it entirely on its own hardware, with its data never leaving the building, and nobody can change or switch off the model under them. The trade-off is honest too — the model card admits foreign models helped prepare parts of the training data, and the specialized nature means it’s built for specific customers, not for everyone. For subscribers watching the open-model race, Kolibri is a credible European answer to the argument that openness and compliance can’t coexist.

FAQ

What is Kolibri?
An open-weight English-German Mixture-of-Experts language model from Germany’s Aleph Alpha, released October 3, 2026, aimed at sovereign on-premise deployment in government and regulated industries.

Where can I get it?
The full weights are available on Hugging Face under the Apache 2.0 license, in BF16 and FP8 formats.

Who is Kolibri for?
Public administration, industrials, and aerospace organizations — and anyone who needs a German-optimized model they can run on their own infrastructure without sending data to a third-party cloud.

Sources: Aleph Alpha (company blog), ForkLog, MarkTechPost.