Aleph Alpha Kolibri: How the sovereign German LLM works

Kolibri is an open-weight large language model from Aleph Alpha for German and English. It is a mixture of experts with 78 billion parameters in total, but it uses only about 3.5 billion of them for each token it reads or writes. It came out on 3 October 2026 under the Apache 2.0 license, with weights on Hugging Face, and was trained from scratch on infrastructure in Germany and Finland.

The author frames Kolibri as an answer to the criticism heard in the US: “you regulate, you don’t innovate.” At SmashingConf New York in 2024, the author wished instead for a middle path: the right balance between innovation and regulation around data privacy, data stewardship, environmental constraints, and energy requirements. Aleph Alpha built Kolibri with the EU AI Act in mind “from the ground up,” and in Aleph Alpha’s own evaluation it scores above every compared model of its size in both German and English. The post draws on Aleph Alpha’s 189-page technical report, the model card, and the launch post, plus one experiment the author ran on its tokenizer; it covers how Kolibri works, where it is strong, where it is not, how to run it, and when it is the right pick.

Kolibri 1 has 78.1 billion parameters in total and 3.46 billion per token, or 4.4% active at a time. It supports German and English, has a native context of 262,144 tokens tested up to 1,048,576 tokens, and is licensed under Apache 2.0 for weights and configuration files, while Aleph Alpha keeps the rights to its training code and methods. It needs about 78 GB of weights in 8-bit floating point, offers four reasoning levels—none, low, medium, and high—supports tool calling, has a knowledge cutoff of 18 June 2026, and was trained on about 24 trillion tokens, more than a fifth of them German, using 768 NVIDIA B200 GPUs.

Aleph Alpha calls Kolibri sovereign, meaning two things. First, it was built in Germany and trained on infrastructure in Germany and Finland under European and German law, with no foreign control. Second, customers get full freedom of deployment and intellectual-property safety, so compliance comes as an inherited property. In plain terms, a ministry or a car supplier can run it on its own servers, with its data never leaving the building, and nobody can change or switch off the model under them. Aleph Alpha has also signed the European Union’s General-Purpose AI Code of Practice.

Sovereign does not mean nothing from outside Europe went into it, and the model card says so: English web text was rephrased with Google’s Gemma 4, German text with Mistral-NeMo, and Qwen3-32B provided labeled data for the quality filters. The training data was then filtered for the political bias such models can have.