Switzerland Enters the AI Arena with Open Source National Model Apertus
The global race for artificial intelligence dominance has a surprising new entrant: an entire nation. Switzerland has officially launched Apertus, a new open-source large language model designed to serve as a public alternative to proprietary models from companies like OpenAI.
Developed by a consortium of the country’s leading public institutions, including the Swiss Federal Institute of Technology in Lausanne (EPFL), ETH Zurich, and the Swiss National Supercomputing Centre (CSCS), Apertus is Latin for open. This name reflects the project’s core philosophy of complete transparency.
Proponents of the model are heralding it as a significant step toward treating AI as essential public infrastructure. Joshua Tan, a leading voice in the movement for public AI, stated that Apertus is the leading public model built by public institutions for the public interest. He called it proof that AI can be a form of public infrastructure like highways, water, or electricity.
A key differentiator for Apertus is its commitment to total openness. The development team has released not only the model itself but also the comprehensive documentation, the full source code of its training process, and the datasets used to create it. This allows researchers, developers, and the public to inspect every aspect of its creation, a level of transparency unheard of with corporate AI models.
The model was also built from the ground up to comply with strict Swiss data protection and copyright laws. This makes it a potentially attractive option for businesses, particularly in regulated industries like finance, that need to adhere to European data sovereignty rules. The Swiss Bankers Association has previously expressed strong interest in a homegrown LLM, citing its potential to better comply with Switzerland’s stringent local data protection and bank secrecy regulations. While Swiss banks already use other AI models, Apertus presents a new, legally compliant option.
Apertus is designed for broad use and is available to researchers, hobbyists, and companies who are encouraged to build upon it and tailor it for specific applications. These could include creating specialized chatbots, translation services, or educational tools.
The model was trained on a massive dataset of 15 trillion tokens spanning over 1,000 different languages. Notably, forty percent of its training data is in languages other than English, including Swiss German and Romansh, highlighting its multilingual capabilities.
In a direct contrast to the practices of some private AI firms, the Swiss team emphasized that Apertus was trained solely on publicly available data. Their web crawlers were programmed to strictly respect machine-readable opt-out requests, commonly found in website files, that ask bots not to scrape content. This approach stands in opposition to AI companies that have faced accusations of ignoring these protocols and scraping copyrighted material without permission, leading to high-profile lawsuits from news organizations and creators.
Apertus is launching in two sizes: an 8 billion parameter model and a more powerful 70 billion parameter version. It is currently accessible for use via the Swiss telecommunications provider Swisscom or through the AI model hosting platform Hugging Face.
The launch of Apertus marks a bold experiment in creating a national, open-source AI infrastructure, challenging the current paradigm of closed, corporate-controlled models and offering a new vision for how the technology can be developed and shared.
