Swiss AI Initiative releases Apertus, a fully open 70B language model trained on Alps

Swiss National Supercomputing Centre (CSCS)/Alps

EPFL, ETH Zurich and CSCS released Apertus in 8B and 70B sizes, trained on 15 trillion tokens covering over 1,000 languages, with weights, data and recipes open. Both versions can be used freely for research, education and commercial work.

• About 40% of the training data is non-English, including Swiss German and Romansh.
• Swisscom deploys it on its sovereign AI platform.
• Training used about 40% of Alps for three to four months, per CSCS director Thomas Schulthess.