Switzerland has released Apertus, a fully open-source AI model designed as a foundational tool for research and development. Developed collaboratively by EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS), Apertus, which means “open” in Latin, offers unrestricted public access to its design, training methodologies, and data. This allows developers and organizations to leverage Apertus in building a diverse range of applications, including sophisticated chatbots, advanced translation services, and innovative educational resources.
Apertus is available in two versions, featuring 8 billion and 70 billion parameters respectively, both licensed under an open-source agreement. This permissive licensing supports a wide range of use cases, including academic research, educational initiatives, and commercial ventures. A key differentiator of Apertus from closed AI systems is its complete transparency; the model’s architecture, training dataset, and comprehensive documentation are all publicly available for scrutiny and modification.
Martin Jaggi, Professor of Machine Learning at EPFL, envisions Apertus as a blueprint for building trustworthy and inclusive AI systems. Thomas Schulthess, Director of CSCS, highlights its potential as a catalyst for innovation. The model’s training involved processing an extensive dataset of 15 trillion tokens across over 1,000 languages, with a significant 40% of the data represented in non-English languages, including Swiss German and Romansh. Imanol Schlag, the project’s technical lead, emphasizes Apertus’s commitment to multilingual capabilities, transparency, and regulatory compliance.
Swisscom is already integrating Apertus into its platform, underscoring its dedication to fostering a secure AI ecosystem. The model can be downloaded from Hugging Face or accessed directly through Swisscom’s services. Developers can gain hands-on experience with Apertus during the Swiss {ai} Weeks event. International users can also access the model via the Public AI Inference Utility, which, according to Joshua Tan, Lead Maintainer, marks Apertus as a leading publicly accessible AI model.
The model’s training adhered strictly to Swiss data protection regulations, copyright laws, and the transparency mandates of the EU AI Act. The dataset comprised only publicly available information, underwent rigorous filtering to eliminate personal data, and was developed in accordance with stringent ethical guidelines. Future developments will concentrate on scaling the model, enhancing its efficiency, and creating specialized tools for applications in law, healthcare, climate science, and education, all while maintaining its core principle of transparency.