Switzerland Launches Multilingual Open-Source AI Model <em>Apertus</em> to Address Transparency Issues

In the intense global race for AI dominance, Switzerland has launched Apertus, its first fully open-source large language model (LLM), developed through a collaboration between ETH Zurich, EPFL, and the Swiss National Supercomputing Centre. The model is designed to be freely accessible to the public, with its code, training data, and technical specifications openly available under an Apache 2.0 license. Amid growing scrutiny over bias, data practices, and legal compliance in major AI systems, Apertus emerges as a trustworthy and transparent alternative.

One of Apertus’ most notable features is its commitment to multilingual inclusivity. Unlike many existing large language models that predominantly use English-based training data, Apertus has been trained on materials covering more than 1,000 languages. This includes not only widely spoken global languages but also less commonly supported ones such as Swiss German and Romansh. Approximately 40% of the training dataset originates from non-English sources, making it more linguistically diverse than many commercial counterparts.

The project became feasible thanks to the Alps supercomputer at the Swiss National Supercomputing Centre in Lugano. The system provided the massive computational power required to train the model while ensuring the process was powered by carbon-neutral energy. Two versions of the model have been released—one with eight billion parameters suitable for lighter applications, and a more powerful 70-billion-parameter variant intended for research and enterprise use.

Ethics and regulatory compliance were central to the development of Apertus. Data collection adhered to website guidelines such as robots.txt to prevent unauthorized scraping of restricted content. The model was built in accordance with Swiss and European regulatory frameworks, including the EU AI Act.

Apertus is already available through multiple channels. Developers and institutions can access it via the Swiss AI platform operated by Swisscom, or directly download it from Hugging Face, the online hub for open-source AI models. To promote experimentation and community engagement, the model will play a central role at Swiss{AI} Week, a series of hackathons and workshops where developers, businesses, and students can test its capabilities and contribute to its evolution.

Importantly, the research team plans to release regular updates to Apertus and develop specialized versions tailored for specific domains such as healthcare, education, law, and climate science. Although it may not match the sheer scale of models like OpenAI’s GPT-5 or Anthropic’s Claude, the creators argue that the significance of a model extends beyond size. According to them, Apertus stands out due to its full transparency, legal compliance, and adaptability—qualities often missing in commercial systems. They also emphasize that Apertus should not be seen as a commercial product, but rather as public infrastructure, akin to roads, electricity, or telecom networks.