Kolibri-1

Kolibri-1 is Aleph Alpha’s open-weight bilingual Mixture-of-Experts reasoning model, designed for efficient German- and English-language AI systems. It contains 78B total parameters while activating only 3.46B per token, providing substantial model capacity with relatively low inference compute. Its 50-layer architecture uses 384 experts per layer, with six routed experts and one shared expert active during processing, alongside a 4:1 combination of sliding-window and grouped-query attention. Kolibri supports explicit reasoning with configurable low, medium, and high effort levels, as well as structured tool calling for agentic workflows. Its native 262K-token context can scale to a validated 1M tokens, enabling long-document processing, RAG, research, and extended coding tasks. The model was pretrained on 20T bilingual and code tokens, followed by additional mid-training and long-context training. FP8 weights reduce its memory footprint to approximately 78 GB.