The llama.cpp project added support for Kolibri-1, a German- and English-language reasoning model from Aleph Alpha, in release b11709. The model has 78 billion parameters, with 3.46 billion active for each token, and uses a mixture-of-experts design.
The update describes support for converting and running Kolibri-1’s model architecture, including its sliding-window attention and FP8 checkpoint format. It does not provide download details or say which devices can run the model.
Why it matters
People using llama.cpp may be able to work with Kolibri-1 through the project, but the release notes do not establish where to get its weights or what hardware is needed.
