Switzerland launches Apertus, an open source, transparent and multilingual AI model
A new LLM that marks the beginning of a journey towards transparent, inclusive and collectively beneficial artificial intelligence.
3' min read
3' min read
Last July, the Swiss Federal Institute of Technology Zurich, the Swiss Federal Institute of Technology Lausanne and the Swiss National Supercomputing Centre announced their collaboration to realise a new LLM or large-scale language model: Apertus. The model is now available to organisations and developers for the development of chatbots, educational tools and translation applications based on artificial intelligence.
Its name, of course, is not accidental as it derives from the Latin word meaning 'open' and reflects precisely its distinguishing characteristic: transparency and free access, for anyone to know its architecture, training data and development documentation. This means that it allows researchers, practitioners but also AI enthusiasts to customise it according to their own needs and objectives, and to analyse each phase of its evolution, unlike the traditional models already known which usually restrict access to selected components.
"Apertus was created for the public good. It is one of the few LLMs on this scale that is completely open and is the first of its kind to embody multilingualism, transparency and conformity as fundamental design principles,' says Imanol Schlag, technical leader of the LLM project and scientific researcher at the ETH Zurich, while it is Martin Jaggi Professor of Machine Learning at EPFL and member of the steering committee of the Swiss AI Initiative, adds 'Our aim with the publication of this first version is to offer a reference model for the development of reliable, sovereign and inclusive AI'.
Apertus is accessible free of charge through its partner Swisscom or can be downloaded from the Hugging Face platform to be implemented in third-party projects and is available in 8 billion parameter versions, more suitable for personal use, and 70 billion parameter versions, both distributed under a permissive open source licence. Trained on 15 trillion tokens and more than 1,000 languages, the model also includes usually neglected idioms such as Swiss German or Romansh for example, never before underrepresented in LLM models. Furthermore, as announced by Swisscom, the model will be constantly updated by the team responsible for its development, which includes specialised engineers and numerous researchers from CSCS, EPFL and ETH.
The upcoming hackathons as part of Swiss {ai} Weeks will offer developers the chance to try out the new model, measure its performance and make useful contributions for future evolutions, but for those who already have technical skills, using Apertus is not complex even if its operation requires additional tools such as servers, cloud infrastructure or dedicated user interfaces: "Swisscom is proud to support the implementation of this pioneering language model on our sovereign Swiss AI platform," comments Daniel Dobos, Head of Research at Swisscom, and continues "As a strategic partner of the Swiss AI Initiative, we support Apertus access during Swiss {ai} Weeks, underlining our commitment to creating a secure and responsible AI ecosystem that serves the public interest and strengthens Switzerland's digital sovereignty."
