Aleph Alpha releases Kolibri-1, a 78B German-English MoE with Apache 2.0 weights
In one sentence On October 3 Aleph Alpha published Kolibri-1 on Hugging Face: a mixture-of-experts model with 78 billion total and 3.46 billion active parameters, bilingual German and English, FP8 weights under Apache 2.0, with 262,144 native context tokens extendable to about 1 million.
Aleph Alpha, the German company positioned on European sovereign AI, released Kolibri-1, its first public model of this generation. The weights can be downloaded from Hugging Face under Apache 2.0, so companies and public bodies can run it on their own servers, even fully offline, without depending on a non-European provider.
The model is large but light to run: out of 78 billion parameters it uses about 3.5 billion per generated word, thanks to its mixture-of-experts architecture. It handles only German and English and works on text only. It was trained from scratch on NVIDIA GPUs in data centers in Germany and Finland.
The FP8 weights take about 78 GB: a single high-end GPU such as an H200 or B200 is enough, or two 80 GB A100 or H100 cards. On the corporate side, Aleph Alpha has signed a binding merger agreement with Canada's Cohere, subject to regulatory approval.
Companies
Aleph Alpha
Tools
Kolibri-1
Tags
Sources