Apertus Launches: Switzerland's Fully Open-Source, Multilingual LLM Breaks New Ground in Transparent AI

Switzerland has entered the global AI arena with a groundbreaking approach to foundation models. Today, EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS) released Apertus—Latin for "open"—a multilingual large language model distinguished by radical transparency throughout its development lifecycle. Unlike proprietary models from Big Tech, Apertus publishes its complete architecture, model weights, training data recipes, and intermediate checkpoints under permissive open-source licenses.

"With this release, we aim to provide a blueprint for how a trustworthy, sovereign, and inclusive AI model can be developed," states Martin Jaggi, EPFL Machine Learning professor and Swiss AI Initiative steering committee member. The project delivers two model sizes: an 8B parameter version for individual use and a 70B variant for complex applications, both available via Hugging Face or Swisscom's sovereign AI platform.

The Open-Source Advantage

Apertus pioneers four critical transparency pillars:
1. Full reproducibility: Training datasets and methodologies are fully documented
2. Commercial-friendly licensing: Permissive use for research and commercial applications
3. Compliance by design: Adherence to Swiss/EU data protection and copyright regulations
4. Ethical data curation: Publicly sourced data filtered against opt-out requests and PII

"Apertus stands among the few fully open LLMs at this scale and is the first to embody multilingualism, transparency, and compliance as foundational principles," emphasizes Imanol Schlag, ETH Zurich research scientist and technical lead.

Multilingual by Design

Trained on 15 trillion tokens across 1,000+ languages, Apertus dedicates 40% of its training data to non-English languages—prioritizing underrepresented tongues like Swiss German and Romansh. This positions it uniquely against monolingual-dominated models:

The Swiss AI Ecosystem

Developed through the Swiss AI Initiative with over 10 million GPU hours on CSCS's "Alps" supercomputer, Apertus represents a new technology transfer philosophy. "We're providing foundational infrastructure to foster innovation across the economy," explains Thomas Schulthess, CSCS Director. Strategic partner Swisscom is already integrating the models into its sovereign AI platform, with business access available immediately.

Caption: Apertus architecture diagram showing multilingual training pipeline (Credit: EPFL/ETH Zurich/CSCS)

The upcoming Swiss {ai} Weeks hackathons will enable hands-on experimentation, while the global Public AI Inference Utility will broaden access. "Currently, Apertus is the leading public AI model built by institutions for public interest—proof that AI can be infrastructure like highways or electricity," notes Joshua Tan, the Utility's Lead Maintainer.

The Road to Trustworthy AI

For EPFL NLP head Antoine Bosselut, this is merely the beginning: "The release is a long-term commitment to open, trustworthy AI foundations for global public good." Future iterations will target domain-specific adaptations in law, healthcare, and climate science while maintaining rigorous transparency standards—setting a compelling precedent for public-interest AI development worldwide.

Source: EPFL News