Kolibri: Aleph Alpha's Sovereign Open-Weight Model

German company Aleph Alpha has unveiled a new language model, Kolibri, choosing a symbolic date for the launch — the Day of German Unity. The weights are available on Hugging Face under the Apache 2.0 license.

Kolibri is a bilingual (German-English) Mixture-of-Experts model with 78B total parameters, of which only about 3B are active per token. It supports context lengths of up to 1M tokens. Its small footprint lets customers run it on-premise without sending internal data to third-party services — a key demand in regulated industries.

A bet on sovereignty

Kolibri is pitched as a tool for "sovereign" mission-critical work: public administration, industrials and aerospace. The team stresses that the model was built in Germany, trained on infrastructure in Germany and Finland, under European and German law, and with no foreign control. It was designed from the ground up with the EU AI Act, the General-Purpose AI Code of Practice and the GDPR in mind, with copyright a particular focus.

German here is not a translation by-product: 21.3% of the training tokens are organic German text, while machine translation accounts for just 6%. The authors note that a corpus translated from English drags along a foreign cultural context — and the model ends up reasoning about German realities as if the world were American (a chancellor, not a president).

An honest "I don't know"

Particular attention went to curbing hallucinations. Using an in-house protocol called Merlin-Arthur, the model is trained to decline to answer when the context lacks the necessary information. For a regulated customer, the company argues, knowing how to say "I don't know" is the difference between a pilot and a deployment.

A pipeline, not one-off runs

Kolibri is the product of an internal "Model Factory." The team first built and validated the whole pipeline on a smaller version, Kolibri Origin (30B parameters), then ran the full model through it. The process is implemented as code: any change triggers an end-to-end mini-run, checkpoints land roughly every hour and are automatically scored against dozens of tests. Over 21 days of pre-training on 768 B200 accelerators, there were 38 unplanned interruptions — about one per 10,000 GPU-hours — and all were handled automatically, without human intervention.

On the reported benchmarks, Kolibri with its 3B active parameters holds its own against models four times larger. Yet the company calls its greatest achievement of the year not the pipeline itself, but a team that has learned to build and ship such models from scratch at high velocity.

Source: Hacker News