ETH Zurich, EPFL and the Swiss National Supercomputing Centre released Apertus in 8-billion- and 70-billion-parameter versions on September 2, 2025. The release makes model weights and documentation of the architecture, training data and methods openly available under a permissive licence allowing research and commercial use.
The models were trained on 15 trillion tokens covering more than 1,000 languages, with 40% of the training material in languages other than English. These include Swiss German and Romansh. Developers can download the models through Hugging Face or access them through services including Swisscom's Swiss AI Platform.
