Apertus: Switzerland presents first open and multilingual AI model

Switzerland wants to make a contribution to its own digital sovereignty with its own open Large Language Model (LLM). Apertus is the name of this multilingual LLM, which was developed by ETH Zurich, EPF Lausanne, and the Swiss National Supercomputing Center (CSCS) in Lugano. It is available in two sizes, the smaller of which is also easy to use for private users.

More than 1,000 languages are represented in the training material, including Swiss German and Romansh, which are obvious languages from a Swiss perspective and have so far been underrepresented in many AI models. The data material comprises around 15 trillion tokens, around 40 percent of which are not in English.

Model complies with EU AI Act

Apertus is said to be at the level of Llama 3, the scientists involved say. For now, it is therefore not causing as much of a stir as the Chinese model Deepseek, for example, which shook up the US-dominated AI economy some time ago.

However, the strength of the model is said to lie in what the Latin name suggests. Apertus stands for openness. Architecture, model weights, intermediate checkpoints, training material, and much more are therefore freely available. The model is the first major model to fulfill the transparency obligations of the EU AI Act. It respects opt-out requests and removes personal data and unwanted content before training, the creators emphasize.

Switzerland wants to expand its AI expertise

One goal can already be ticked off with the existence of the model; it is the expansion of AI expertise in research, society and business. Of course, this is only a start, but according to the publishers, Apertus is a kind of foundation that can now be built on. In September, a hackathon at the Swiss AI Weeks will provide the first taste of what is possible with the model.

Apertus can be downloaded from the Huggingface platform, among others. It is available in versions with 8 and 70 billion parameters.

(mki)