Skip to content
Predictive Systems
PSI Daily

Aleph Alpha Releases Kolibri, an Open-Weight German AI Model Trained to Say "I Don't Know"

SovereigntyAllan C. Tan, MS

A German model you can download, run on your own servers, and trust to admit when the answer is not in your documents. Sovereignty just became a file.

A civil servant at dusk reads a floating AI panel that answers "I don't know. This isn't in your documents." beside an on-premise server icon.

A German city office wants an AI assistant to read its case files. It cannot send those files to a cloud in another country, and it cannot afford a model that guesses. On October 3, German Unity Day, Heidelberg's Aleph Alpha released a model built for that office, and anyone can download it.

Kolibri, German for hummingbird, is free on Hugging Face under the Apache 2.0 license. Aleph Alpha says its teams built it in Germany and trained it on infrastructure in Germany and Finland, with the EU AI Act, the EU's code of practice for general-purpose AI and the GDPR in mind. It also published the training data summary that the AI Act requires.

Big model, light touch

Kolibri holds 78 billion parameters but uses only about 3.5 billion per token, or word piece. That design, called a mixture of experts, keeps it fast and cheap to run. It is still not a laptop model. The weights need about 78 GB of memory, so the minimum is one H200 chip or two A100s. It natively reads 262,144 tokens, and Aleph Alpha says it tested about one million. Users can also choose how hard it thinks, from none to high.

Aleph Alpha trained Kolibri with a game it calls Merlin-Arthur, where one player hides the evidence to bait the model into guessing. The winning move is to say "I don't know." On one hallucination test, Kolibri declined instead of answering wrong on 44 percent of the questions it missed, up from 15 percent for its predecessor. These scores are the company's own.

That is the same checkability we look for in System 1 decision models. A model that shows when it is unsure tells you where to look.

One open question: Aleph Alpha has signed a deal to combine with Canada's Cohere, pending approval.

Sovereignty used to be a promise. Now it is a download.

Sources

  1. Aleph Alpha, "Kolibri Has Landed: A Sovereign Open-Weight Model," Oct. 3, 2026
  2. Hugging Face, "Aleph-Alpha/Kolibri-1," accessed Oct. 4, 2026
  3. Aleph Alpha, "Public Summary of Training Content for General-Purpose AI Models," Oct. 3, 2026
  4. RuntimeWire, "Aleph Alpha releases German AI model with 78 GB of FP8 weights," Oct. 3, 2026
  5. WirtschaftsWoche, "Aleph Alpha Kolibri: Neues KI-Sprachmodell für Verwaltung," Oct. 3, 2026
  6. Reuters, "Cohere, Aleph Alpha combine to target enterprise AI market," Sept. 16, 2026