Switzerland has unveiled Apertus, an entirely open-source AI model developed by EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS), available now on HuggingFace. Aptly named after the Latin term for “open,” the model stands as a publicly transparent alternative to closed systems like OpenAI’s ChatGPT or Anthropic’s Claude
Open by Design: Full Access Provided
Apertus offers complete transparency—its source code, model weights, documentation, training methodology, and datasets are freely accessible under the permissive Apache 2.0 license. This approach embraces innovation by enabling users to inspect, adapt, or build upon every aspect of the model.
Language Inclusivity and Ethical Data Practices
With support for over 1,000 languages—including underrepresented languages like Swiss German and Romansh—and containing around 15 trillion training tokens (40% non-English), Apertus is designed for global linguistic equity and relevance. Crucially, its development strictly used publicly available data, honoring websites’ machine-readable “opt-out” directives and avoiding unethical data practices like stealth crawling.
Powered by Switzerland’s “Alps” Supercomputer
Apertus was trainied on Alps, one of the world’s most powerful, carbon-neutral supercomputers located at CSCS. Equipped with over 10,000 NVIDIA Grace Hopper GPUs, Alps enabled efficient large-scale model training on Swiss infrastructure—reinforcing AI sovereignty and ecological responsibility.
Catalyst for Innovation and Sovereign AI
Apertus is more than a model; it’s the centerpiece of Switzerland’s broader Swiss AI Initiative, aimed at promoting trustworthy, sovereign, and inclusive AI systems. It’s expected to be a base for applications in research, education, government, and industry, offering affordably compliant infrastructure.
