Aleph Alpha released its English- and German-language reasoning model Kolibri under Apache 2.0. Its mixture-of-experts architecture activates about 3.46 billion of roughly 78 billion parameters per token and supports tool calling and a context window of up to 1,048,576 tokens. Trained in Germany and Finland, it provides weights for organizations to deploy in their own environments. The official model card recommends no more than 262,144 tokens for efficiency and complex tasks and says the FP8 weights alone occupy about 78GB. Low active computation should not be confused with suitability for small devices.
Back to Tech
Aleph Alpha releases Kolibri weights for English and German reasoning
Few active parameters do not remove the memory needed for all weights; complex tasks are advised to stay within 262,144 tokens.