Aleph Alpha, the Heidelberg AI company, released Kolibri on Saturday, Oct. 3, the Day of German Reunification. It is an open-weight model: its trained files are published, so anyone can download it and run it on their own computers. The files are on Hugging Face under the Apache 2.0 license, which lets anyone use, change and share them free of charge. Aleph Alpha says the license does not cover its code or the methods used to train the model.
Aleph Alpha says it built Kolibri for German and English work in regulated areas such as public administration, industry and aerospace, and that it is small enough to run on a customer's own servers without sending internal data to outside services. Kolibri has 78 billion parameters, the numbers a model tunes as it learns. But it is a mixture-of-experts model, a design that switches on only part of the model for each piece of text, so it runs faster than its full size suggests: about 3.5 billion parameters work at a time.
It still needs about 78 GB of memory and, at minimum, two of Nvidia's data-center AI chips, either the A100 (80 GB version) or the H100, or one H200, B200 or B300. It can take in about 262,000 tokens at once, or about 1 million with extra settings. Tokens are the small chunks of text, often parts of words, that AI models read and write, so 262,000 tokens is roughly 200,000 English words. Aleph Alpha recommends the lower limit when speed matters and for complex tasks.
In Aleph Alpha's own tests, Kolibri beat three open rivals from the spring on math and science: Alibaba's Qwen3.6-35B-A3B, Nvidia's Nemotron 3 Super and Mistral Small 4. It scored 96.9 on the AIME 2025 math test and 84.3 on the GPQA Diamond science test. It trailed Qwen3.6 on the BFCL v4 test of using software tools (61.4 against 67.2) and the LongBench Pro long-document test (64.5 against 70.8). Aleph Alpha's own tables also show Alibaba's newer Qwen3.8 27B, a model that uses all 27 billion of its parameters at once, ahead overall in both languages: 80.2 against 75.5 in English and 79.9 against 70.8 in German.
The comparisons have limits. The main charts set Kolibri against rivals from the spring, and newer open models such as Z.ai's GLM-5.3 and Moonshot AI's Kimi K3 are missing from the tables. As of Saturday, no outside group had ranked Kolibri against other models. Aleph Alpha says it trained the model to answer "I don't know" when the documents it is given lack the answer. Aleph Alpha also reports results on a public test of hallucinations, cases of a model stating something false as if it were true. On the questions Kolibri did not get right, it declined to answer, rather than answer wrong, 44 percent of the time, up from 15 percent for its unreleased predecessor. Qwen3.6 did so 56.7 percent of the time.
It is not yet clear how Kolibri does in tests run by others. Nor is it clear how its scores on Aleph Alpha's internal tests, built for sectors such as the car industry and the German public sector, carry over to real work. Aleph Alpha has also agreed to merge with the Canadian AI company Cohere, and the combined company is to operate under the Cohere name, so Kolibri's longer-term future is uncertain.
Correction (Oct. 4): an earlier version said Kolibri declined to answer, rather than answering wrong, 44 percent of the time on a public hallucination test. That rate covers only the questions it did not answer correctly.