# What is Kolibri-1, and what does a "sovereign" open model really give you? Date: 2026-10-04 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/kolibri-1-aleph-alpha-sovereign-model/ **Answer:** Kolibri-1 is an English-German open-weight model from Germany's Aleph Alpha, released on October 3, 2026, with 78 billion parameters and about 3.46 billion active per token. It is not a top scorer on general tests; its pitch is local control, strong German and a model that fits on one H200 or B200 chip. Germany's Aleph Alpha released Kolibri yesterday, on the day of German Reunification. The date was no accident. The company describes it as a specialized model for sovereign, mission-critical work in public administration, industry and aerospace. **Sovereign** here means the customer keeps control of where the model runs and whose laws apply, instead of sending data to a service abroad. Canadian readers have a reason to watch too. Aleph Alpha has signed a binding agreement to merge with Toronto's Cohere, according to Trending Topics, with the combined company taking the Cohere name and Heidelberg remaining a research hub. So this model may matter here as well. ## What was released | Item | Kolibri-1 | |---|---| | Total parameters | 78 billion | | Active parameters per token | 3.46 billion | | Languages | German and English | | Context | Up to 1,048,576 tokens; 262,144 or less recommended | | Memory for the weights | About 78 GB at FP8 | | Files | About 79 GB (FP8) or 156 GB (BF16) | | Licence | Apache 2.0 | A **parameter** is one learned number, and a **token** is a small piece of text. Kolibri is a **mixture-of-experts** model: 384 expert blocks per layer, of which 6 are chosen for each token, plus one always in use. This keeps compute low while memory stays high, which the card states plainly as the trade-off. The company's blog rounds the active figure to 3 billion. ## How it works Most of the model's choices follow from focusing on two languages. The pre-training corpus was about 62.5 percent English, 23.9 percent German and 13.6 percent code, 20 trillion tokens in total, on 768 NVIDIA B200 chips over 21 days. Aleph Alpha also built its own **tokenizer**, the component that cuts text into tokens. German makes long compound words, and Trending Topics gives the example of "Bundessozialgerichtes", which Kolibri splits into four meaningful parts. Long documents get special handling too. Most attention layers look only at nearby text, and a smaller number look across the whole context, in a four-to-one ratio. Think of reading a long contract with attention on the current page, and occasional checks back through the whole file. Trending Topics also reports that the model was trained to say it does not know when the answer isn't in the documents, which is valuable for government files, but it needs checking on your own. ## What it takes to run This is where Kolibri stands out, because you can do the planning from one line of the card: | Setup | Status on the card | |---|---| | One B300, one B200 or one H200; or two H100 SXM5; or two A100 cards of 80 GB | Minimum | | One B300 or one B200; or two H200; or two H100 SXM5 | Recommended | That is one server, not a cluster, and far closer to what a ministry or a mid-size company can buy or rent. At about 78 GB, an ordinary gaming card won't hold it. Community conversions may fit bigger desktops, but quality and speed then need checking. The card says Kolibri requires Aleph Alpha's own inference package, which provides a vLLM plugin, or its container image, so setup takes more than a single click. The card also recommends human review. It says the model belongs on the advisory side of decisions, not as the component that decides, which fits how a public body should use it in any case. ## The benchmark debate Aleph Alpha compares Kolibri with Mistral Small 4, NVIDIA's Nemotron 3 Super and Alibaba's Qwen3.6-35B-A3B. It leads several tests in that group, with 84.3 percent on GPQA Diamond and 85.9 percent on LiveCodeBench, and Qwen3.6 leads on agent tasks and long context. Trending Topics argues that this field flatters the model, since those rivals are months old. On an overall index, it counts 20 or more open-weight models ahead of Kolibri, and it says models of similar active size, such as Alibaba's Qwen3.8-Flash-Next, are well ahead. Aleph Alpha replies that public benchmarks reflect its customers poorly, and has built its own tests for public administration, automotive suppliers and aerospace, where it reports clear gains over its earlier internal model, Kolibri Origin. Both points can be true. A German agency scoring models on German administrative tasks may get a different ranking from a global index. The honest way to decide is to run your own test, with your own documents. > **Try it:** If you work in German or with German clients, take ten real questions about a document you know, then run them on Kolibri through a host and on one general model. Count the correct answers and the times each model admits it isn't sure. ## What does Kolibri's Apache 2.0 licence actually cover? The weights and configuration files are under Apache 2.0, which allows commercial use, modification and redistribution with no revenue threshold. You need to keep the licence and notices and mark changed files. The card adds a limit that is easy to miss. It says the rights apply only to the weights and configuration files in the repository. Code, architecture, parameter settings and training methods stay outside the grant, and Aleph Alpha keeps its rights to those. For someone running or fine-tuning the model, this changes little. For someone hoping to copy the training pipeline, it matters, and it's a good reminder that **open weights** are not the same as a fully open process. So "sovereign" has limits too. Kolibri was trained on NVIDIA chips, the serving tools come from a wider ecosystem, and a merger could change who steers future versions. What you do control, once the files are on your server, is where your data goes. For many public bodies, that is the part that matters most. ## Sources - Kolibri-1 model card: https://huggingface.co/Aleph-Alpha/Kolibri-1 - Kolibri has landed: https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/ - Aleph Alpha's Sovereign A.I. Model Kolibri Is No Match for the Open-Weight Leaders: https://www.trendingtopics.eu/aleph-alpha-kolibri-open-weight/ --- From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0 (https://creativecommons.org/licenses/by-nc-sa/4.0/): credit the guide, link to https://guidetoartificialintelligence.com/en/learn/kolibri-1-aleph-alpha-sovereign-model/, no commercial use without permission, and keep the same licence.