Aleph Alpha released Kolibri on October 3, making the weights of its English-German language model available under Apache 2.0. The model targets reasoning and agent tasks, with particular emphasis on German-language work.
Kolibri has 78 billion parameters in total and activates about 3 billion for each token. This mixture-of-experts design uses a subset of the model at a time, reducing the computation involved in generating an answer. The complete weights still have to be stored and made available to the runtime.
Aleph Alpha advertises support for context lengths up to one million tokens. Its own evaluations describe the model’s language and reasoning performance; the context limit alone does not show how accurately it will retrieve a particular detail from a long document.