In early September 2025, ETH Zurich, EPFL and the Swiss National Supercomputing Centre presented Apertus, a large language model built as part of the Swiss AI Initiative. Like ChatGPT, it can generate text, answer questions or translate between languages. The difference is that Apertus is not a commercial product. It is fully open: the code, the data used for training and the full pipeline that make the model work are all published and documented. The researchers describe it as the first step toward treating AI as public infrastructure, comparable to roads, schools or libraries.
Large language models learn from very large collections of text and then predict the next word to form an answer. Apertus was trained on more than fifteen trillion text fragments. About forty percent of that material is not in English. It covers over one thousand languages, including Swiss German and Romansh, reflecting an effort to make the system less dependent on English and more inclusive.
Apertus can already be tested through the Public AI platform, where users interact with it in a simple chat interface. The broader aim, however, is not just to provide another chatbot. By releasing every component, the project invites universities, companies and public services to build trustworthy tools for education, health or administration, and to adapt them to local needs and laws.
The importance of this openness is that independent experts can examine how the model was built, check for bias and design safeguards. Institutions can fine tune it with their own documents without depending on a single vendor. This helps align AI with transparency, accountability and linguistic inclusion. It also reduces reliance on proprietary systems that may restrict access or prioritise commercial interests.
Yet Apertus is not a panacea. Like any large model, it can make mistakes, reflect stereotypes, and its largest version requires substantial computing power and energy. This means that most people will not run it themselves but will access it through third-party providers. The difference is that Apertus makes this process transparent. Its strengths and weaknesses can be studied, discussed and improved openly. If researchers, companies and public bodies continue to refine it, Apertus could become a form of shared digital infrastructure. In that role, it shows how powerful AI can be developed in the open and adapted for the benefit of society.
Researcher in Energy Systems at ISAAC – SUPSI