Another new artificial intelligence? Yes—and it’s Swiss! Its name? Apertus, a large language model (LLM) developed by a consortium of public institutions:
- École Polytechnique Fédérale de Lausanne (EPFL),
- ETH Zurich,
- The Swiss National Supercomputing Centre (CSCS).
This project aims to redefine AI standards by putting openness, transparency, and the common good at the heart of its development.
An AI designed for the public good
Apertus, whose name means “open” in Latin, embodies a vision: making AI a public infrastructure, accessible to everyone.
Unlike giants like ChatGPT, this model isn’t trying to compete on raw performance, but to offer a reliable and ethical alternative. Comparable in capabilities to Meta’s Llama 3 model released last year, Apertus stands out for being fully open source.
Its creators have made available not only the model weights, but also the training source code and complete documentation, all under the Apache 2.0 license, accessible via the Hugging Face platform.
This choice of openness serves one goal: enabling organizations—whether businesses, academic institutions, or governments—to adapt the AI to their specific needs. Available in two versions, one with 8 billion parameters and the other with 70 billion, Apertus can be deployed locally, ensuring that processed data stays on the user’s servers.
This feature makes it especially attractive for sensitive sectors like finance, healthcare, or law, where data confidentiality is crucial.
Multilingualism and ethics at the heart of the project
Apertus training is based on a massive corpus of 15 trillion tokens, collected from public sources on the Internet.
Notably, 40% of this data comes from non-English content, covering more than 1,000 languages. This multilingual approach, built in from the start, aims to make AI accessible to a wide range of people and cultural contexts.
The researchers also took care to comply with regulations, including the European Union’s AI Act, by excluding personal data and undesirable content, and by avoiding sources that explicitly prohibit their use for AI training.
Here’s the presentation video for this new AI:
An alternative to the AI giants?
While Apertus doesn’t yet match the performance of the most advanced proprietary models like GPT-5 or Grok4, its accessibility and transparency make it a unique asset. By allowing users to customize the model and run it locally, it meets a growing need for secure and adaptable solutions.
The teams behind Apertus are already working on improved versions, further strengthening its potential for real-world applications.
This Swiss project, led by public institutions, shows that it’s possible to combine technological performance, transparency, and regulatory compliance. With immediate availability on Hugging Face and ambitions to evolve, Apertus could well become a benchmark for organizations looking for reliable, sovereign AI in Europe.
Source: EPFL
