Swiss research institutions have collaboratively introduced a groundbreaking open-source AI model intended to serve as a foundational platform for future innovations and applications. Developed by EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS), this model is named Apertus-derived from Latin, meaning “open.” This title embodies the project’s commitment to full transparency, with every aspect of its architecture and training process openly accessible to the public.
Introducing Apertus: A New Era of Open AI
Apertus is designed to empower developers and organizations to build a variety of AI-driven solutions, including conversational agents, multilingual translation systems, and educational tools. The model is available in two configurations: an 8-billion-parameter version and a more robust 70-billion-parameter variant. Both editions are distributed under a permissive open-source license, enabling unrestricted use in academic research, commercial ventures, and educational initiatives. Users can download the model directly or access it via Swisscom, a key strategic partner in this initiative.
Commitment to Full Transparency and Sovereignty
Unlike many proprietary AI models that limit disclosure, Apertus offers complete openness regarding its design, training datasets, and comprehensive documentation. Martin Jaggi, Professor of Machine Learning at EPFL and a member of the Swiss AI Initiative Steering Committee, emphasized that Apertus serves as a blueprint for developing AI systems that are trustworthy, sovereign, and inclusive. The model will receive continuous updates from a dedicated team of engineers and researchers across the collaborating institutions.
Thomas Schulthess, Director of CSCS and Professor at ETH Zurich, highlighted Apertus as a catalyst for innovation, aiming to bolster AI expertise across academia, industry, and society. He noted that this project transcends typical technology transfers by establishing a sustainable infrastructure for long-term AI development and deployment.
Extensive Multilingual Capabilities
The training corpus for Apertus encompasses an impressive 15 trillion tokens spanning over 1,000 languages, with approximately 40% of the data sourced from non-English content. This includes underrepresented languages such as Swiss German and Romansh, which are often excluded from large language models. Imanol Schlag, the project’s technical lead and Research Scientist at ETH Zurich, remarked that Apertus is among the few large-scale, fully open language models that prioritize multilingualism, transparency, and regulatory compliance as foundational principles.
Swisscom has integrated Apertus into its sovereign AI platform, reinforcing its dedication to fostering a secure, responsible AI ecosystem that upholds Switzerland’s digital sovereignty. Daniel Dobos, Research Director at Swisscom, underscored the importance of this deployment in serving the public interest.
Accessing and Utilizing Apertus in Real-World Scenarios
While downloading Apertus is straightforward for technically proficient users, practical implementation requires adequate computational infrastructure such as servers or cloud platforms. During the ongoing Swiss {ai} Weeks, which run through October 5, 2025, developers can experiment with Apertus via a Swisscom-hosted interface, including participants in hackathons. Swisscom’s business clients already have access to the model through its AI platform, and international users will soon be able to utilize Apertus through the Public AI Inference Utility.
Joshua Tan, Lead Maintainer of the Public AI Inference Utility, described Apertus as the foremost publicly available AI model developed by public institutions for the collective benefit. He likened it to essential public infrastructure-comparable to utilities like highways, water, or electricity-highlighting its role in democratizing AI access.
Ensuring Ethical Standards and Legal Compliance
All components of Apertus-including training data, model weights, and intermediate checkpoints-are openly shared under an open-source license. The training adhered strictly to Swiss data protection laws, copyright regulations, and the transparency mandates outlined in the EU AI Act. The dataset was carefully curated to include only publicly available information, with rigorous filtering to exclude personal data and respect website opt-out requests. Ethical considerations guided the removal of inappropriate content prior to training.
Looking Ahead: The Evolution of Apertus
Antoine Bosselut, Professor at EPFL and Co-Lead of the Swiss AI Initiative, stated that Apertus exemplifies how generative AI can be both powerful and openly accessible. He emphasized that the model’s release marks the start of a sustained effort to develop open, reliable, and sovereign AI technologies that benefit the global community.
Future plans include expanding the Apertus model family, enhancing computational efficiency, and creating specialized tools tailored to sectors such as law, healthcare, climate science, and education-all while maintaining rigorous transparency and ethical standards.
