Apertus, a fully open large language model developed by the Swiss National AI Initiative, represents one of the most ambitious attempts yet to build a foundation model that is transparent by design, compliant with regulation from the ground up, and deliberately independent of the commercial incentives that drive models from US-based labs.
The project is a collaboration between EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS). Its name is Latin for “open,” and its core claim is that every component — training data, model weights, source code, methodology, and alignment principles — is documented, reproducible, and publicly available.
Apertus is released in two sizes: 8 billion parameters and 70 billion parameters, each available in Base and Instruct variants optimised for chat. The models were trained on 15 trillion tokens of text drawn from openly available data, covering more than 1,800 languages, with approximately 40% of the training corpus allocated to non-English content.
The multilingual emphasis is deliberate. Most large language models are trained primarily on English, with other languages treated as secondary. Apertus includes underrepresented languages such as Swiss German and Romansh, as well as hundreds of languages that have seen minimal inclusion in other foundation models. The team behind Apertus argues that a model trained on the world’s linguistic diversity cannot be built by a lab optimising for English-language benchmarks and commercial deployment.
Compliance as a design principle
Apertus was designed to meet the requirements of the European Union’s AI Act from the start — not as a retrofit, but as a structural constraint on how data is collected, filtered, and used. The training corpus respects machine-readable opt-out requests from websites, even retroactively. Personal data is filtered before training begins. The model uses the “Goldfish objective” during pretraining to suppress verbatim memorisation of training data while maintaining performance.
This approach distinguishes Apertus from many open-weight models, which are released with permissive licences but whose training data cannot be fully audited. As the EU AI Act enters enforcement phases, models whose data provenance is opaque face legal risk in European markets. Apertus’s developers argue that compliance is a feature, not a limitation, and that it positions the model for adoption by governments and regulated industries that cannot risk using black-box AI systems.
European AI sovereignty in practice
Apertus is not a research curiosity. Swisscom, the country’s largest telecommunications provider, is a strategic partner and offers the model on its sovereign Swiss AI Platform, keeping data and inference within Swiss jurisdiction. The model is also available on Hugging Face and through the Public AI network.
The geopolitical timing matters. The US government’s recent export control action against Anthropic’s Fable 5 and Mythos 5 models, and the broader uncertainty around access to American AI infrastructure, has accelerated interest in European-developed alternatives. Apertus offers a template: a model built by a publicly funded academic consortium, trained on EuroHPC supercomputing resources, compliant with European regulation, and available under a permissive open-source licence that permits commercial use.
“With this release, we aim to provide a blueprint for how a trustworthy, sovereign, and inclusive AI model can be developed,” Martin Jaggi, professor of machine learning at EPFL and a member of the Swiss AI Initiative steering committee, said at launch.
Peer review and future plans
Apertus’s technical report was accepted at the ACL 2026 main conference, one of the top venues in natural language processing, and is available on arXiv. The paper presents the model’s architecture, training methodology, and safety evaluation.
The team plans to expand the Apertus model family with future versions that improve efficiency, explore domain-specific adaptations for fields such as law, climate, and health, and integrate additional capabilities while maintaining the project’s commitment to transparency. Apertus Mini, a set of 16 smaller models demonstrating distillation and quantisation techniques, was released in June 2026.
The bigger picture
Apertus sits at the intersection of several trends that define the current AI landscape: the push toward open models as a counterweight to proprietary labs, the growing regulatory pressure from the EU AI Act, and the scramble by governments outside the US and China to secure independent AI capabilities.
Whether any single model can satisfy all of these demands — openness, compliance, multilingual breadth, and competitive performance — remains an open question. Apertus’s early benchmarks place it in the same range as other open models at equivalent scale, but it does not claim to match the frontier capabilities of GPT-5.5 or Claude Opus 4 on English-language tasks.
What Apertus offers instead is a demonstration that the choice between “powerful” and “open” and “compliant” may be false. The project’s backers are betting that for governments, regulated industries, and anyone who cannot afford legal ambiguity around how their AI was built, a fully documented, regulation-ready model is more valuable than one that is merely state-of-the-art.
Sources: Apertus official site (2026); Apertus ACL 2026 paper; Europesays (May 23); Apertus AI Wiki