Aleph Alpha releases Kolibri, Germany's sovereign open-weight LLM

German AI company Aleph Alpha released Kolibri, an open-weight large language model for German and English, on 3 October 2026. The weights are on Hugging Face under the Apache 2.0 license, and it was trained from scratch on infrastructure in Germany and Finland. Aleph Alpha calls it sovereign: a company or public authority can run it on its own servers, its data never leaves the building, and nobody outside can change or switch off the model.

What Kolibri is

Kolibri is a mixture of experts (MoE) with 78.1 billion parameters, with only about 3.5 billion (4.4%) active per token. Each of its 50 layers holds 384 experts plus one shared expert, and a router sends every token to just 6 of the 384. It was trained on roughly 24 trillion tokens, more than a fifth of them German, on 768 NVIDIA B200 GPUs, with a knowledge cutoff of 18 June 2026.

The context window is 262,144 tokens natively, and Aleph Alpha tested it up to 1,048,576. The weights take about 78 GB in 8-bit floating point, so it needs data-center hardware: two 80 GB NVIDIA A100 or H100 GPUs, or one H200, B200 or B300. It supports tool calling and four reasoning levels, and its UniBPE tokenizer needs 11.2% fewer tokens for German than GPT-5's.

Why it counts as sovereign

Aleph Alpha says German teams built it and trained it on German and Finnish infrastructure under European and German law, with no foreign control. The open weights give customers freedom of deployment and IP safety, and the company signed the EU's GPAI Code of Practice.

Sovereign does not mean nothing from outside Europe went in, and its own model card says so: English web text was rephrased with Google's Gemma 4, German with Mistral-NeMo, and Qwen3-32B provided labeled data for the quality filters. The training data was also filtered for political bias.

Strong points and weak ones

In Aleph Alpha's own evaluation, Kolibri scores above every compared model of its size in both languages: 75.5 overall in English and 70.8 in German. It is strongest at math, with 87.5 on AIME 2025 in German and 96.9 in English, and it reasons in German on German prompts instead of switching to English. It was also trained to abstain: when the evidence in the context does not support an answer, it says so, which it did 44% of the time in a test where the answer was unknown, against 23.7% for GPT-OSS 120B.

The weak spots are public too: it ranks last of the 12 compared models on a closed-book test, trails its rivals on multi-turn tool calling and coding tasks, and its 78 GB of weights fit only data-center GPUs. The vLLM plugin is new, and no hosted provider serves the model yet.

Kolibri is the right pick when German text and controlled hardware both matter: an authority, a bank or a manufacturer keeping documents in house that would rather hear "I don't know" than a confident wrong answer. Retrieval over long German documents such as laws, contracts and manuals plays to its biggest strengths at once.