Aleph Alpha puts a 78B sovereign model on open weights

On October 3, Aleph Alpha announced Kolibri, a German-English mixture-of-experts transformer with 78 billion total parameters and 3 billion active parameters. The company says it supports up to a 1 million-token context window and released weights on Hugging Face under Apache 2.0.

For makers, the combination of German focus, a small active footprint, long context, and a permissive license gives teams a local model to test where data control matters. Aleph Alpha positions Kolibri for reasoning, mathematics, agentic work, and regulated sectors; its benchmark figures are company-reported, not independent proof.

Try the weights with a narrow German or long-document workload, then measure quality, memory use, and throughput on your own hardware. Watch the model card, evaluation code, and community replications before treating the headline parameter count as an operating advantage.