Switzerland launches its own open-source AI model

There’s a new player in the AI race, and it’s a whole country. Switzerland has just released Apertus, its open-source national Large Language Model (LLM) that it hopes would be an alternative to models offered by companies like OpenAI. Apertus, Latin for the world “open,” was developed by the Swiss Federal Technology Institute of Lausanne (EPFL), ETH Zurich and the Swiss National Supercomputing Centre (CSCS), all of which are public institutions.

“Currently, Apertus is the leading public AI model: a model built by public institutions, for the public interest. It is our best proof yet that AI can be a form of public infrastructure like highways, water, or electricity,” said Joshua Tan, a leading proponent in making AI a public infrastructure.

The Swiss institutions designed Apertus to be completely open, allowing users to inspect any part of its training process. In addition to the model itself, they released comprehensive documentation and source code of its training process, as well as the datasets they used. They built Apertus to comply with Swiss data protection and copyright laws, which makes it perhaps one of the better choices for companies that want to adhere to European regulations. The Swiss Bankers Association previously said that a homegrown LLM would have “great long-term potential,” since it will be able to better comply with Switzerland’s strict local data protection and bank secrecy rules. At the moment, Swiss banks are already using other AI models for their needs, so it remains to be seen whether they’ll switch to Apertus.

Anybody can use the new model: Researchers, hobbyists and even companies are welcome to build upon it and to tailor it for their needs. They can use it to create chatbots, translators and even educational or training tools, for instance. Apertus was trained on 15 trillion tokens across more than 1,000 languages, with 40 percent of the data in languages other than English, including Swiss German and Romansh. Switzerland’s announcement says the model was only trained on publicly available data, and its crawlers respected machine-readable opt-out requests when they came across them on websites. To note, AI companies like Perplexity have previously been accused of scraping websites and bypassing protocols meant to block their crawlers. Some AI companies have also been sued by news organizations and creatives for using their content to train their models without permission.

Apertus is currently available in two sizes with 8 billion and 70 billion parameters. It’s currently available via Swisscom, a Swiss information and communication technology company, or via Hugging Face.