On 3 October 2026, the Heidelberg AI company Aleph Alpha released the language model Kolibri. The weights are available on Hugging Face under the permissive Apache 2.0 licence. Companies can therefore run the model on their own infrastructure without sending data to an external provider. Kolibri specialises in German and English and targets regulated sectors such as public administration, industry and aerospace. It is not designed for a laptop or a small office server, though.
What is Kolibri?
Kolibri is a so-called mixture-of-experts model with 78 billion parameters in total. Only around 3.5 billion of them are active per word, which makes it faster and cheaper to run than a classic model of the same size. Further key facts according to Aleph Alpha:
- Languages: German and English; more than a fifth of the training data is German.
- Context: up to one million tokens, meaning very long documents at once.
- Capabilities: step-by-step reasoning at four levels and tool use, for example for agents.
- Training: on infrastructure in Germany and Finland.
- Knowledge cut-off: June 2026.
How good is Kolibri?
According to Aleph Alpha, Kolibri matches or beats models with up to four times as many active parameters in maths, coding and long documents. In general knowledge and multi-step tool use, however, it trails comparable models such as Qwen. Important: all figures come from the manufacturer itself; independent tests are still pending.
What should you view critically?
- Open, but not entirely: the licence covers the weights and configuration, not the code, architecture or training methods.
- Own software stack: running it requires the aleph-alpha-inference package, an extension for the widely used vLLM software.
- "Sovereign" with caveats: according to heise, Aleph Alpha also used other providers' models to generate synthetic training data.
- Company future: in September 2026, Aleph Alpha agreed to merge with the Canadian provider Cohere, subject to approval. What happens to Kolibri is open.
What hardware does Kolibri need?
Kolibri is built for data-center graphics cards. According to Aleph Alpha, it handles 18 simultaneous requests with long documents on two Nvidia H100 cards. For a mid-sized company, that means its own GPU server in a data center or rented GPU capacity from a German provider. For smaller tasks there are lighter open models that run on normal hardware, more in Local AI in practice.
Who is Kolibri worth it for?
For companies that process many German documents, may not hand over their data and have, or want to rent, the infrastructure for a model of this size. Anyone wanting to experiment first is better off starting with a smaller model. An overview of other European models is given in Sovereign AI from Europe.
