Swiss AI Model Apertus: An Open Multilingual Model

In September 2025, EPFL, ETH Zurich, and CSCS introduced a new language model called Apertus. This model is Switzerland's first large open multilingual system, emphasizing transparency and diversity in generative artificial intelligence. The name Apertus comes from the Latin word for "open," which accurately captures its essence—the entire development process, including the architecture, model weights, training data, and recipes, is fully available and documented. Developers and organizations can use it as a foundation for their own applications, such as chatbots, translation systems, or educational tools.

The Apertus model was developed as part of the Swiss AI Initiative, led by EPFL and ETH Zurich. Researchers, engineers, and students from across Switzerland collaborated on it, together with CSCS and its infrastructure. Funding was provided through an investment of more than 10 million GPU hours on CSCS's Alps supercomputer and support from the ETH Board, which oversees the ETH Domain comprising ETH Zurich, EPFL, WSL, PSI, Empa, and Eawag. Strategic partners also contributed, particularly Swisscom, Switzerland's largest telecommunications provider.

Technical Details and Training

Apertus is available in two sizes: one with 8 billion parameters and another with 70 billion parameters. The smaller version is better suited to individual use, while the larger one offers more robust capabilities. Training was conducted on 15 trillion tokens across more than 1,000 languages, with 40% of the data being non-English. Supported languages include Swiss German, Romansh, and many others that have previously been overlooked in similar models.

The developers ensured compliance with Swiss data protection and copyright laws, as well as transparency obligations under the EU AI Act. The training data comes exclusively from publicly available sources that were filtered to honor website opt-out requests, including retroactively. Personal data and undesirable content were removed before the training itself. This approach ensures that the model is ethical and reliable for a variety of applications.

According to researcher Imanol Schlag, the project's technical lead and a scientist at ETH Zurich, Apertus was built for the public good. It is one of the few fully open models of this size to treat multilingualism, transparency, and regulatory compliance as fundamental principles. Comparisons with Meta's Llama 3 model underscore its competitiveness, even though it focuses more on openness than on directly competing with proprietary systems such as ChatGPT or Claude.

Availability and Use

Users can access Apertus through strategic partner Swisscom or download it from the Hugging Face platform, where it has been released under a permissive open license. This license allows its use in education, research, and commercial applications. For people outside Switzerland, the model is also available through the Public AI Inference Utility, led by Joshua Tan. According to him, Apertus represents a leading public AI model created by public institutions in the public interest, much like infrastructure such as roads or electricity.

Setting up Apertus is straightforward for professionals and experienced users, but it requires additional components such as servers, cloud infrastructure, or user interfaces. The first opportunity for testing is offered by Swiss {ai} Weeks, a series of hackathons and events where developers can experiment with the model and provide feedback. Swisscom offers a dedicated interface for participants and gives its corporate customers access through its Swiss AI platform.

Extensive resources were released alongside the model: documentation, the source code for the training process, datasets, and model weights including intermediate checkpoints—all under an open license. The terms are available on Hugging Face.

Future Plans

The development team, which includes specialized engineers and researchers from CSCS, ETH Zurich, and EPFL, plans to update the model regularly. According to Martin Jaggi, professor of machine learning at EPFL and a member of the Swiss AI Initiative steering committee, Apertus is intended to serve as a model for trustworthy, sovereign, and inclusive AI. Thomas Schulthess, director of CSCS and professor at ETH Zurich, sees Apertus as a tool for strengthening AI expertise in research, society, and industry.

Daniel Dobos, research director at Swisscom, emphasizes the company's pride in deploying this model on the Swiss AI platform, thereby supporting a secure and responsible ecosystem. Antoine Bosselut, professor and head of the Natural Language Processing Laboratory at EPFL, describes the release of Apertus as the beginning of a journey toward open, trustworthy, and sovereign AI for the public good worldwide.

Future versions are expected to expand the model family, improve efficiency, and add specializations for fields such as law, climate, health, and education, while maintaining high standards of transparency. The Swiss {ai} Weeks initiative, supported by the EPFL AI Center, ETH AI Center, Swisscom, Impact Hub Switzerland, Kickstart Innovation, Panter, and Wildcard Media, brings together researchers, developers, entrepreneurs, and citizens in an effort to advance ethical and open AI.

Source: ethz.ch