Aleph Alpha's Kolibri: Germany's Sovereign AI Model

Kolibri is an open-weight large language model (LLM) developed by Aleph Alpha, designed for both German and English. Released under the Apache 2.0 license, it features 78 billion parameters but efficiently uses only about 3.5 billion for each token. The model was trained in Germany and Finland, adhering to European and German laws, and is positioned as a sovereign AI solution, emphasizing data privacy and compliance with the EU AI Act. Kolibri’s architecture includes a mixture of experts, a specialized tokenizer for long German words, and sliding-window attention to manage long contexts. It is optimized for German reasoning and designed to minimize hallucinations by training it to admit when it doesn’t know.