Launch of Open AI Model Apertus by Swiss Institutions

A group of Swiss institutions has announced the launch of a new open artificial intelligence model named Apertus, designed to serve as a foundation for future applications and research. This model was developed by the École Polytechnique Fédérale de Lausanne (EPFL), the Swiss Federal Institute of Technology in Zurich, and the Swiss National Supercomputing Centre (CSCS). The name, which means “open” in Latin, reflects the core principle of the model, as every part of its design and training process is available to the public.

What is the Apertus Model?

Developers and organizations can use Apertus to create chatbots, translation tools, or educational applications. The model can be downloaded directly from Hugging Face or accessed through Swisscom, the strategic partner of the initiative. It is available in two versions: a model with eight billion parameters and a larger one with 70 billion parameters. Both are available under an open-source license that permits their use in research, education, and commercial projects.

Apertus distinguishes itself from other AI systems by being a fully open model, with its architecture, training data, and documentation accessible to all for examination.

Model Goals and Guidelines

The model aims to provide a blueprint for a reliable, sovereign, and inclusive AI model. According to Professor Martin Jaggi, a member of the steering committee of the Swiss AI Initiative, Apertus will be regularly updated by a team of engineers and researchers from CSCS, the Swiss Federal Institute of Technology in Zurich, and EPFL.

Thomas Schulthess, Director of CSCS and Professor at the Swiss Federal Institute of Technology in Zurich, described Apertus as “an engine for innovation and a means to enhance AI expertise in research, society, and industry.”

Linguistic Diversity in Apertus

The training process included 15 trillion tokens in over 1,000 languages, with about 40% of the data in non-English languages. The model includes languages often neglected in large language models, such as Swiss German and Romansh.

This reflects the initiative’s commitment to providing a multilingual, transparent, and legally compliant AI model.

Transparency and Compliance

Under the open-source license, the training data, model weights, and intermediate checkpoints are available. The training process adhered to Swiss data protection rules, Swiss copyright law, and transparency requirements according to the European Union’s AI regulations.

It was essential to filter publicly available data to remove personal information and respect opt-out requests from websites. Ethical guidelines were also applied to exclude undesirable material before training commenced.

Conclusion

The Apertus model demonstrates that generative AI can be both powerful and open. The launch of Apertus marks the beginning of a long journey towards providing open, reliable, and sovereign AI foundations for the benefit of the public worldwide. The teams responsible for the model look forward to developing it further to include improvements in efficiency and domain-specific tools such as law, health, climate, and education, while continuing to uphold strict transparency standards.