Apertus: Switzerland’s Open, Multilingual Answer to the AI Infrastructure Question
On September 2, 2025, researchers from ETH Zurich, EPFL and the Swiss National Supercomputing Centre (CSCS) released Apertus, Switzerland’s first large-scale, fully open multilingual language model. The name comes from the Latin word for “open,” and the project treats that idea as an engineering requirement rather than a marketing label. Its architecture, weights, training methods and data documentation are published so that other researchers and organisations can inspect, reproduce and extend the system. 1 2
The launch placed a public research project at the centre of a question usually dominated by private technology companies: who should be able to understand, adapt and govern the foundation models that increasingly shape digital services?
A model built for inspection
Many AI releases make model weights available while keeping other parts of the process difficult to audit. Apertus takes a broader approach. The project describes its release as including the model weights, the training data sources, data-preparation code, training recipes, evaluation material and intermediate checkpoints. The original v1 family was offered in 8-billion- and 70-billion-parameter versions under the Apache 2.0 licence, which permits research, education and commercial use subject to the licence terms. 1 2
That distinction is important. A researcher can download weights, but reproducible science requires more than a file containing parameters. It also requires enough information to understand what entered the training pipeline, how the data was filtered, how the model was evaluated and which choices shaped its behaviour.
The technical report presents Apertus as an attempt to address two weaknesses in the open-model ecosystem: limited multilingual representation and weak visibility into data rights and preparation. It reports pretraining on about 15 trillion tokens, with roughly 40 percent of the material in non-English content. The report describes coverage of more than 1,800 languages, while the launch communication highlighted more than 1,000 languages and specific attention to Swiss German and Romansh. The difference reflects the way language inventories are counted across datasets and evaluation settings, not a contradiction about the project’s multilingual ambition. 1 3
Swiss languages and a wider linguistic remit
Switzerland has four national languages: German, French, Italian and Romansh. Yet a model designed only around the largest global languages would still miss much of the country’s linguistic reality. Apertus gives space to Swiss German and Romansh while also targeting languages that receive limited representation in commercial systems.
This matters for more than translation. Language coverage affects public information, education, cultural archives, research tools and the ability of smaller communities to develop applications in their own linguistic settings. A model can be technically impressive in English and still be poorly suited to a multilingual public institution.
The project’s broader language strategy also connects Switzerland to a global problem. Most training data and evaluation resources remain concentrated in a relatively small group of languages. Giving more languages a place in a public model does not automatically solve bias or quality gaps, but it creates an inspectable base on which local researchers can measure progress and build specialised systems.
Openness linked to compliance
Apertus was designed around publicly available data and documented filtering steps. The project says it respected machine-readable opt-out requests, including retroactive requests where possible, and filtered for personal information and other unwanted material before training. The technical report also describes the use of a training objective intended to reduce verbatim memorisation while preserving useful performance. 1 3
These measures should not be confused with a guarantee that a large language model is free from errors, bias or legal uncertainty. Openness makes scrutiny easier; it does not remove the need for scrutiny. Data provenance, copyright claims, privacy requests, benchmark design and model behaviour still require continuing review.
The value of the approach is that those questions can be asked against a visible record. Developers can examine the stated sources and methods rather than relying only on assurances from a model provider. Regulators, civil-society groups and researchers can also test the system against the principles described by its creators.
From academic project to public infrastructure
The Swiss AI Initiative, co-led by ETH Zurich and EPFL, provides the institutional setting for Apertus. The initiative began in December 2023 and combines research, engineering and access to CSCS’s Alps supercomputer. Its public materials describe an open-science programme intended to release software, models and data artifacts that can be used by startups, small and medium-sized companies, large organisations and the public sector. 2 4
The launch was supported by more than 10 million GPU hours on Alps, alongside funding from the ETH Domain and contributions from strategic partners such as Swisscom. Access was offered through Swisscom’s Swiss AI platform, Hugging Face and the Public AI network, giving users different routes to experiment with the model or deploy it in their own environments. 1
That distribution model reveals a practical point: a permissive licence does not make a 70-billion-parameter system effortless to run. Users still need suitable hardware, storage, software frameworks and operational expertise. The 8-billion-parameter version is more approachable for individual users, but serious deployment remains an infrastructure project.
What changed after the first release?
Apertus did not stop at the original launch. The project’s current roadmap lists smaller Apertus 1.1 models for mobile and browser deployment, followed by Apertus 1.5, an 8B and 70B family with image and audio understanding, a longer context window, improved instruction following, tool use and an optional reasoning mode. These later releases extend the original open-model philosophy rather than replacing the significance of the first public release. 2 5
The 1.5 documentation states that the models accept text, images and audio as input and produce text as output. It also reports support for contexts of up to 262,144 tokens. Such capabilities broaden the possible applications, but they add new evaluation questions: how reliable is image interpretation across cultures, how are audio accents represented, and how transparent are the additional training stages?
The answers will depend on continued documentation and independent testing. The strongest part of Apertus is therefore not a single benchmark score. It is the attempt to make the entire model lifecycle available for examination while still delivering a system large enough to be useful.
Closing Thoughts
Apertus is most interesting when viewed as a choice about institutional design. Switzerland did not simply announce another chatbot. Its research organisations released a model together with a public record of how that model was assembled.
The approach has clear limits. Transparency cannot guarantee factual accuracy, eliminate harmful outputs or settle every dispute around training data. Open weights can also be difficult to operate safely when users lack the expertise to evaluate them. Yet those limitations are easier to confront when the underlying process is visible.
For multilingual societies, that visibility has special value. A public model can be tested in languages and domains that commercial evaluation may overlook. It can also give local universities, companies and public bodies more control over where sensitive workloads run and how adaptations are made.
What is the true implication of this?
The true implication is that foundation models can be treated as shared research infrastructure rather than only as proprietary services. That does not mean every organisation should train its own model. It means that more organisations can inspect the foundation, challenge its assumptions, adapt it to local needs and participate in deciding what responsible development looks like.
Apertus also demonstrates that multilingual AI is a governance issue. Language inclusion depends on data choices, evaluation priorities and access to computing resources. A model that exposes those choices gives communities a better chance to influence them.
What makes this significant?
Apertus brings together three commitments that are often separated: broad language coverage, open technical artefacts and attention to data compliance. Its importance will ultimately be measured by what others can verify, improve and build with it.
If the project maintains that standard through future releases, it may offer a durable reference point for public-interest AI: not a claim that open systems are perfect, but a practical demonstration that powerful language technology can be developed with more accountability and wider participation.
References
Comments