Switzerland has launched Apertus, an open-source large language model developed by national research institutions and supercomputing centers, emphasizing transparency, public access, and compliance with the EU AI Act.

Apertus is available for download on Huggingface in versions with 8 and 70 billion parameters, designed to be competitive with commercial models while prioritizing transparency and legal compliance.

The model is trained on 15 trillion tokens across over 1,000 languages, including underrepresented dialects like Swiss German and Romansh, using only open-source data respecting site restrictions.

Apertus can be run on local servers, giving users control over their data, which is especially appealing for sectors like finance, health, education, and climate research.

The project aims to address issues such as AI hallucinations, biases, and copyright infringement, creating a safer, more controlled AI environment aligned with European standards.

Legal and data privacy aspects are managed through transparency documents, with clear processes for copyright and privacy requests, ensuring compliance with regulations.

Switzerland envisions Apertus as a form of public infrastructure to serve the public interest, reducing reliance on private corporations and enhancing data privacy, especially for Swiss banks.

Future plans include regular updates, expanding the model family, and developing domain-specific versions for sectors like law, climate, health, and education, while maintaining transparency and responsible AI principles.

Apertus supports long context processing up to 65,536 tokens and can be deployed via APIs compatible with OpenAI, facilitating integration into various applications.

The development is part of a broader Swiss AI initiative involving over 800 researchers, funded by the Swiss government and strategic partners, with over 10 million GPU hours invested.

The project emphasizes fostering trustworthy, inclusive AI, with ongoing updates from a dedicated team of researchers and engineers, aiming to serve societal and economic impacts.

While early reactions are positive, challenges remain in establishing governance to prevent misuse and ensuring responsible deployment, given the community oversight model.