Switzerland has released a multilingual open-source large language model known as Apertus. Developed by EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS), the model is not intended as a commercial product. It functions instead as a public digital infrastructure initiative.
Apertus is built around the principles of transparency, inclusivity, and accessibility. Every part of the model—from its training data to intermediate development checkpoints—is openly available. Unlike proprietary AI models offered by tech corporations, Apertus positions itself as an alternative grounded in legal compliance and open research.
Transparency as a Starting Point
Where most models offer limited insight into how they’re trained or what data they ingest, Apertus lays everything out. The source code is public. The training documentation is detailed. The model weights, including intermediate versions, are freely accessible.
The dataset used includes only publicly available content. No private data scraping occurred, and all machine-readable opt-out signals from websites were respected. The website may also ask for removal from the training dataset after the fact. Such granular consent tracking is rare, which speaks to growing concerns around ethical AI development.
With data protection regulations tightening worldwide, this approach presents a practical guide to brands and developers assessing third-party models.
A Global, Multilingual Model
Apertus was trained on over 15 trillion tokens in more than 1,000 languages. Crucially, over 40% of this data was sourced from non-English languages. These include regionally specific languages such as Swiss German and Romansh, as well as a broader selection of under-represented linguistic groups.
This multilingual foundation makes the model especially relevant for regions where English is not the primary language. Organisations working in Africa, Asia, Eastern Europe, or Latin America often face limitations with language models trained predominantly on English. Apertus offers an opportunity to build language-sensitive applications at scale.
Scalable for Different Needs
Two versions of Apertus are available: an 8-billion-parameter model for lightweight tasks and a 70-billion-parameter version for more computationally demanding applications. This range provides flexibility depending on infrastructure capabilities and use-case scenarios.
The models are hosted across multiple platforms. They can be downloaded from Hugging Face, deployed through Swisscom’s Sovereign Swiss AI Platform, or accessed via a public AI inference utility. This removes traditional barriers to AI deployment—licensing, vendor lock-in, and opaque APIs.
Built with Compliance at the Core
Apertus meets legal standards set by Swiss and European regulatory bodies. This includes full alignment with the General Data Protection Regulation (GDPR), strict data provenance protocols, and responsible sourcing of training materials.
The training process avoids any personal data and strictly adheres to copyright laws. Some commercial models, which do not disclose their training sources, cannot be audited for development under those circumstances.
So, in these respects, the other product is not an option for institutions working with sensitive or regulated data, such as financial services, healthcare, government, and education.
Practical Applications in Global Markets
Apertus lends itself to practical deployment across industries. Its multilingual capabilities are especially relevant in education, public administration, translation services, and customer support operations.
Governments can use the model to power citizen services. Universities may integrate it into multilingual research environments. Media outlets can experiment with content localisation. Legal firms may explore compliant document analysis workflows.
This wide range of use cases stems from the model’s foundational transparency and accessibility.
Open Access Without Geographic or Institutional Barriers
The model is publicly freely accessible. Whether in Jakarta, Nairobi, Buenos Aires, or Toronto, institutions and individuals can freely and fully download and deploy Apertus.
Such open access stands in sharp contrast to many tools in AI that are restricted with respect to geography, amount of cases in a licence tier, and so on. Apertus has been conceived from the perspective that AI serving the public interest should be freely available – especially in low-resource environments.
Such positioning aligns with ongoing global conversations around AI equity and inclusion, particularly in the Global South.
A Public Model in a Commercial World
In terms of scale, Apertus matches commercial offerings like Meta’s LLaMA 3. But it exists outside commercial competition. The purpose is not to capture market share but to offer a publicly governed alternative.
This is aligned with Switzerland’s broader strategy around digital sovereignty. Apertus serves as an infrastructure layer, not a profit centre. It is built to be audited, trusted, and improved upon by the global research and development community.
The intention is clear: Apertus is a reference point for how public AI can be done.
Evaluation and Integration for Organisations
For enterprises considering integration, Apertus presents a low-friction starting point. The smaller model size can be tested in pilot workflows. Documentation is readily available. Deployment options include both local installations and managed services.
Besides being useful for compliance teams as the instrument to audit AI tools, the model also offers technical teams a concrete case study of scalable multilingual AI architecture.
Even if not adopted immediately, Apertus can set certain internal benchmarks. What should transparency look like in AI? How can multilingual performance be measured fairly? These are questions every AI-using organisation now faces.
A Model Worth Watching
The bigger backdrop for Apertus is a global movement towards making AI development open, accountable, and accessible to all worldwide. Whether this approach might gain traction in other parts of the world or influence commercial actors to rethink their strategy remains to be seen.
What is evident is that Apertus shifts expectations. It sets a precedent for what can be offered in the public interest. And it does so without compromising technical scale or multilingual depth.
For organisations building AI strategies, especially those operating internationally, this model deserves a place on the roadmap.