Aleph Alpha Releases Kolibri-1: Open-Source 78B MoE Model with 1M Context Window
German AI lab Aleph Alpha has released Kolibri-1, an open-weights Mixture-of-Experts (MoE) language model featuring 78 billion total parameters and 3.46 billion active parameters per token. The model is distributed under the permissive Apache 2.0 license and supports context lengths of up to 1 million tokens. By activating only 3.46B parameters during inference, Kolibri-1 dramatically lowers memory requirements and hardware costs while providing extensive long-context processing. It offers European and global developers a powerful, sovereignty-focused open model optimized for both English and German applications. Kolibri-1 incorporates explicit reasoning modes and native tool-calling capabilities alongside its sparse routing architecture. Its small active parameter footprint enables significantly faster text generation speeds and higher throughput per GPU compared to traditional dense models of similar total capacity.
## BACKGROUND
Mixture-of-Experts (MoE) is an architecture that routes input tokens to specific dynamic subsets of parameters called experts, rather than computing across the entire network for every token. This approach enables LLMs to scale their knowledge capacity without proportionally increasing the computational power needed during inference.