A consortium of Swiss research institutes has introduced Apertus, a family of language models designed as a transparent, privacy-focused alternative to commercial systems like ChatGPT.
Apertus was developed by the Swiss AI Initiative, a partnership between ETH Zurich, EPFL, and Switzerland’s National Supercomputing Centre (CSCS). The models are available in 8- and 70-billion parameter versions and are positioned as an alternative to US and Chinese tech companies' offerings.
The goal of Apertus is not to compete with the massive budgets of leading AI firms. Instead, the focus is on providing a secure and accessible AI system for both scientific and business applications. The development team aims to set a standard for building trustworthy, sovereign, and inclusive AI models.
Fully open, not just open-weight
Apertus stands out for its full transparency. Unlike many "open-weight" models that release only the final weights, the Swiss team is publishing every artifact from the development cycle, including data preparation scripts, training code, evaluation tools, and intermediate checkpoints. This approach is intended to make independent review and further development possible.
Strict data compliance was a priority throughout the project. Apertus models were trained exclusively on publicly available data, with full respect for robots.txt exclusions. This policy was enforced retroactively: opt-out preferences as of January 2025 were applied to all prior web scrapes to protect content owners’ rights. The data was filtered to remove copyrighted, non-permissive, toxic, and personally identifiable material, supporting compliance with the EU AI Act.
To reduce the risk of the model memorizing training data, the team used the "Goldfish objective," a method that selectively masks tokens during training and limits rote memorization.
Technical foundation and swiss values
Apertus was trained on the "Alps" supercomputer at CSCS, using up to 4,096 Nvidia GPUs and a dataset containing 15 trillion tokens. The project prioritized multilingual coverage, with training data spanning more than 1,800 languages and roughly 40 percent sourced from non-English material. Swiss national languages, including Romansh and Swiss German, were specifically included in the dataset.
For model alignment on sensitive topics, the team applied principles from the "Swiss AI Charter," which is based on Swiss constitutional values such as neutrality and consensus-building. They used a separate language model as an "LLM-as-judge" to check Apertus’s responses for alignment with these values.
Benchmarking shows that the Apertus-70B-Instruct model performs reliably but generally falls behind leading open-weight models in most broad categories. For tasks involving knowledge retrieval and logical reasoning, Apertus-70B usually scores lower than established models like Llama-3.3-70B, Qwen2.5-72B, or OLMo-2-32B-Instruct. The difference is especially clear in complex reasoning tasks, reflecting that Apertus is not specifically built for these scenarios.
However, Apertus’s specialized strengths are most evident in multilingual applications. The technical report notes that Apertus-70B-Instruct consistently outperforms Llama-3.3-70B when translating between German and six Romansh language variants.
Potential for swiss industry
Swiss industry groups see strong potential in a domestically developed AI model, especially for meeting local privacy and banking requirements. The Swiss Bankers Association points to long-term benefits, while some large banks, such as UBS, are already using solutions from OpenAI and Microsoft.
Swissmem, the mechanical engineering association, notes the advantages of a model built for European data regulations, though broad industry adoption is not guaranteed. In practice, international solutions may still be the best fit for some needs.
Apertus is recognized in the open-source community as a significant step forward for open models, largely due to its scale and the computing power involved in its development. The models are available for researchers, businesses, and the public on Hugging Face and can be tested on PublicAI.
