Summary:
1. Swiss institutions release a new open AI model called Apertus, designed for research and applications.
2. Apertus is fully open, multilingual, and transparent, allowing for various uses and deployment options.
3. The model follows strict standards of transparency, compliance, and ethical guidelines to serve the public interest.
Article:
A groundbreaking development in the world of AI has emerged, as a collaboration between EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS) has produced a new open AI model named Apertus. This model, which translates to “open” in Latin, sets itself apart by allowing public access to every aspect of its design and training process. Developers and organizations now have the opportunity to leverage Apertus for creating chatbots, translation tools, and education-focused applications. Available in two versions – an 8-billion-parameter model and a larger 70-billion-parameter version – Apertus is released under a permissive open-source license, enabling its use in research, education, and commercial projects.
Unlike many other AI systems that keep certain details under wraps, Apertus is characterized by its complete openness. The model’s architecture, training data, and documentation are readily accessible for scrutiny by the public. This commitment to transparency is underscored by the words of Martin Jaggi, Professor of Machine Learning at EPFL, who emphasizes the aim of providing a blueprint for the development of a trustworthy, sovereign, and inclusive AI model. Regular updates to Apertus will be carried out by a dedicated team of engineers and researchers from CSCS, ETH Zurich, and EPFL.
One of the standout features of Apertus is its multilingual reach, achieved through a training process that involved a vast amount of data in over 1,000 languages. Notably, the model incorporates languages often overlooked by Language Model Models (LLMs), such as Swiss German and Romansh. Imanol Schlag, technical lead of the project and Research Scientist at ETH Zurich, describes Apertus as a model built for the public good, emphasizing its rare combination of multilingualism, transparency, and compliance as foundational design principles.
To ensure the model’s accessibility and real-world usability, Apertus can be tested through various means, including the ongoing Swiss {ai} Weeks and the Swisscom-hosted interface for hackathon participants. Swisscom business customers can already begin utilizing the model through the company’s AI platform, with international users gaining access through the Public AI Inference Utility. The commitment to transparency and compliance is further reinforced by the adherence to Swiss data protection rules, Swiss copyright law, and the transparency requirements of the EU AI Act during the model’s training process.
Looking ahead, Apertus represents more than just a technological advancement – it symbolizes a commitment to open, trustworthy, and sovereign AI foundations for the public good worldwide. Antoine Bosselut, Professor at EPFL and Co-Lead of the Swiss AI Initiative, expresses the sentiment that the release of Apertus marks the beginning of a journey rather than a final step. Future updates will focus on expanding the model family, enhancing efficiency, and developing domain-specific tools for various sectors like law, health, climate, and education, all while upholding the highest standards of transparency.
