Apertus: Open Foundation Model for Sovereign AI

Overview of Apertus

Apertus is a fully open foundation model designed to enable sovereign AI development. Developed by the Swiss AI Initiative—a collaboration between EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS)—Apertus provides open weights, open data, and open science to ensure transparency and reproducibility in AI training.

Openness and Reproducibility

Apertus distinguishes itself from many "open weights" models by providing full transparency into the training process. The project provides open access to the training data, code, weights, methods, and alignment principles. By documenting and making these components reproducible, Apertus aims to be the AI equivalent of Open Source software.

Regulatory Compliance and Data Privacy

Apertus is engineered to meet the requirements of the EU AI Act. To ensure compliance and safety at scale, the model incorporates the following data handling practices:

• PII Removal: The model removes personally identifiable information (PII) from its training set.

• Memorization Prevention: The model is built to prevent the memorization of training data, reducing the risk of data leakage.

• Opt-out Respect: The model respects data opt-outs, ensuring that data owners have more control over their training sets.

Performance and Multilingual Capabilities

Apertus is available in scales of 8B and 70B parameters, which are competitive with other top open models of equivalent size. The model is designed for multilingualism from the start, having been trained on over 1,000 languages, enabling it to provide a broad global foundation for AI applications.

Strategic Partnerships

The development of the Swiss AI Initiative is the Swisscom same as a Strategic Partner, providing the infrastructure and support necessary to build a sovereign AI foundation for the European region.