Swiss Initiative Unveils Major Open-Source AI Model

Swiss institutions have launched Apertus, one of Europe’s most extensive fully open language models, offering a transparent alternative to proprietary AI platforms. Available in two configurations—8 billion and 70 billion parameters—it is comparable in scale to Meta’s Llama 3 from 2024, yet distinguishes itself through open design, public availability, and lawful compliance. Swiss Initiative Unveils Major Open-Source AI Model, a fitting restatement of the headline, underscores this ambition.

Researchers from EPFL, ETH Zurich and the Swiss National Supercomputing Centre designed Apertus to embody trust and inclusivity, building on publicly available datasets only, while scrubbing personal or unauthorised content and honouring machine-readable exclusion directives. The development was supported by over 10 million GPU hours on CSCS’s “Alps” supercomputer and funding from the ETH Board and strategic partners such as Swisscom.

The dataset covers 15 trillion tokens spanning more than 1,000 languages, with about 40 per cent non-English content—including Swiss German, Romansh and other under-represented languages. This multilingual orientation reinforces the model’s commitment to serving diverse linguistic communities.

Apertus is openly documented from cover to cover: its architecture, model weights, training recipes, intermediate checkpoints, and data sources are all available under a permissive licence on platforms such as Hugging Face. This level of openness extends to compliance with Swiss data protection rules and EU AI transparency obligations.

Swisscom is enabling access via its sovereign AI platform and promoting hands-on experimentation during the Swiss {ai} Weeks events, while the Public AI Inference Utility offers global access. Researchers emphasise that Apertus is not a concluding milestone but rather the foundation of ongoing advancement in open, sovereign AI.