Aleph Alpha's Kolibri is an open model built to run on your own hardware

German AI company Aleph Alpha released Kolibri, an open-weight model aimed at regulated users who need to run everything on their own hardware. It is a Mixture-of-Experts transformer with 78.1 billion parameters in total but only 3.46 billion active per token, drawn from 384 experts with six active at a time. That design is the point. Aleph Alpha says it matches models with up to four times the active parameter count on reasoning tasks while staying small enough to run on-premise, so a public agency or defense contractor never has to send internal data to an outside inference service.

The numbers are competitive. On AIME 2025 it scores 96.9 in English and 87.5 in German; on GPQA Diamond, 84.3 and 81.3; on HumanEval+, 92.7 percent. It was pre-trained on 20 trillion tokens (62 percent English, 21 percent German, 14 percent code) over 21 days on 768 B200 GPUs, and handles context up to one million tokens. The weights are on Hugging Face under Apache 2.0. Aleph Alpha also trained the model to abstain when the context does not support an answer, and in the announcement it stresses a transparent supply chain and training done in Germany and Finland under European law.

Why it matters

If you work in European public administration or another regulated sector, a bilingual model you can download and run in your own datacenter removes the usual reason to avoid frontier AI: that prompts would leave your network.