Madiha Yaqoob
German AI company Aleph Alpha published Kolibri on October 3, a 78-billion-parameter language model with its full weights downloadable from Hugging Face under the Apache 2.0 license, meaning anyone can download, run, modify, and commercially use it without paying Aleph Alpha a license fee or routing requests through its API.
A model built to run on someone else’s servers
Per Aleph Alpha’s own blog post announcing the release, Kolibri is a mixture-of-experts model with 78.1 billion total parameters but only 3.46 billion active per token, which keeps inference costs down relative to its size, and it supports context windows up to one million tokens. It was trained on 768 Nvidia B200 GPUs running on German and Finnish infrastructure, specifically so that government agencies, defense contractors, and regulated industries could deploy it entirely on their own hardware under European law, with no dependency on a foreign cloud provider.
Sovereignty as the selling point
Independent technical coverage of the release notes Kolibri’s architecture deliberately trades some raw benchmark performance for predictable, auditable behavior: 384 experts with only six active per token, full attention in just 10 of its 50 layers, and a sliding window on the rest to contain compute costs. It is not pitched as a GPT-6.1 Sol or Claude Sonnet 5.5 competitor on raw capability. It is pitched at organizations that are legally or politically unwilling to send their data to an American lab’s servers at all.
What this opens up for people outside that use case
Most freelancers and small agencies have no sovereignty requirement to satisfy, but Kolibri’s license terms apply to them just as much as to a German ministry. A free, commercially-licensed model of this size is one more option for anyone trying to build an AI-powered product or service without an ongoing per-token API bill, provided they can get their hands on hardware capable of running a 78-billion-parameter model, which is the real constraint for most individual developers rather than the license itself. The model card and full weights are live on Hugging Face now for anyone who wants to try.
