Apertus is the leading Swiss language model: a fully open large language model (LLM), released on 2 September 2025 by the Swiss Federal Institute of Technology in Lausanne (EPFL), the Swiss Federal Institute of Technology in Zurich (ETH Zurich) and the Swiss National Supercomputing Centre (CSCS). Updated to version 1.5 on 24 July 2026, Apertus already translates documents for the Ticino cantonal administration and assists a Basel newsroom. Alongside it, other Swiss projects target narrower uses: SwissBERT at the University of Zurich, MeditronFO for medicine, and the models of the Lausanne start-up Giotto.

What is Apertus, the Swiss language model?

Apertus is a foundation model, not a consumer application. In its announcement of 2 September 2025, EPFL presents it as a building block for creating assistants, translation systems or educational tools.

Its openness is what sets it apart. The architecture, the model weights, the data and the training recipes are all published and documented. Many so-called open models release only the weights, with no way of checking how they were made.

The model comes in two sizes: 8 billion and 70 billion parameters. According to EPFL, the smaller one is better suited to individual use. Both are released under the Apache 2.0 licence, which also permits commercial use, according to the CSCS announcement of 24 July 2026.

The technical model card published on Hugging Face, the platform for distributing AI models, details the method. Apertus was trained from scratch, rather than derived from an existing model, on 15,000 billion tokens of web text, code and mathematics. It claims native support for 1,811 languages. EPFL specifies that 40% of the data is not in English, including languages that are scarce elsewhere, such as Swiss German and Romansh.

Who is developing Apertus, and with what resources?

Apertus is the product of the Swiss AI Initiative, led by EPFL and ETH Zurich. According to the initiative's website, it got under way in December 2023 with more than 10 million hours of computing time on graphics processing units (GPUs) and a grant of 20 million francs from the ETH Domain. It is the first initiative of the Swiss National AI Institute, a partnership between the ETH AI Center and the EPFL AI Center.

The project is led by Martin Jaggi and Antoine Bosselut, professors at EPFL's School of Computer and Communication Sciences, and by Imanol Schlag, a researcher at ETH Zurich, CSCS specifies. Training ran on Alps, the CSCS supercomputer in Lugano, using 4,096 GH200 chips according to the model card. Swisscom, the country's largest telecommunications operator, is the strategic partner and contributes to the funding, according to EPFL.

The legal framework was considered from the outset. According to EPFL, the model takes into account Swiss data protection and copyright law, as well as the EU AI Act. The corpus, limited to public data, respects websites' opt-outs from data collection, even retroactively.

Key figures

  • 15,000 billion training tokens, 40% of them in languages other than English.
  • 20 million francs: the ETH Domain's initial grant to the Swiss AI Initiative.
  • More than 800 researchers, including 70 professors specialising in AI, from more than 10 institutions.
  • More than 10,000 GH200 chips in the CSCS Alps supercomputer.
  • 262,144 tokens of context for Apertus 1.5, four times more than the first version, according to Liip.

What has changed with Apertus 1.5?

Apertus 1.5 is an update, not a new model. "Rather than introducing an entirely new model, Apertus 1.5 incorporates feedback from the first version and adds several capabilities," explains Martin Jaggi in the CSCS announcement. The model now understands images and audio as well as text. It reasons better, follows instructions better and is better at calling external tools.

The company Liip, which runs Apertus in its ZüriCityGPT assistant, notes in a blog post of 3 September 2026 that the context window grows from 65,536 to 262,144 tokens. The model can thus read much longer documents in one go. An optional "thinking" mode has been added.

The same release launched Apertus Mini, a suite of 16 compact models designed for modest hardware. CSCS is also opening an inference service, in other words access through an application programming interface (API), for its users and the Swiss academic community.

What is Apertus already used for in Switzerland?

CSCS documents three uses of the first version. The canton of Ticino uses it for an internal translation service, which avoids handing sensitive administrative documents to a commercial provider. The Basel online outlet Bajour hosts it locally to prepare the political news for its daily newsletter Basel Briefing.

Bajour has also built a search tool for the minutes of the Grand Council, offered to other newsrooms through the We.Publish publishing ecosystem. "With Apertus, we know what we are building on," says Samuel Hufschmid, an editor at Bajour. The third use, in medicine, is still at the research stage.

What other Swiss language models exist?

Apertus is not the only project, but the others do not share its ambition. The table positions them by developer, type and use, based on their respective publications.

ProjectDeveloperTypeTerms of useIntended use
Apertus 1.5Swiss AI Initiative (EPFL, ETH Zurich, CSCS)General-purpose generative model, 8 and 70 billion parametersApache 2.0, commercial use permittedFoundation for assistants, translation, business tools
MeditronFOEPFL's LiGHT laboratoryMethod for specialising an open model in medicineData, code and procedures publishedMedical research, clinical trials in preparation
SwissBERTDepartment of Computational Linguistics, University of ZurichText analysis model, 153 million parametersCommercial use excludedWord tagging, text classification and vector representations, including Romansh
Giotto Solo and AltoGiotto, a Lausanne start-upReasoning modelsLicence not specified in the announcementChat service, API, on-premises installation

MeditronFO, released by EPFL on 17 June 2026, is a production pipeline. It turns an open model into a medical version, drawing on more than 46,000 clinical practice guidelines. The LiGHT laboratory applied it to OLMo, EuroLLM and Apertus. The Apertus 70B version gained 6.6 points on medical exams compared with the base model. This result concerns exams, not patients: the team is preparing clinical trials, named MED.USE, from Switzerland to Tanzania.

SwissBERT belongs to a different family. This University of Zurich model was trained on more than 21 million Swiss news articles, published up to the end of 2022, in Standard German, French, Italian and Grisons Romansh. It does not write text: its authors intend it for analysis, such as word tagging, text classification or vector representations of sentences.

Giotto, finally, is a company. As our article on Giotto's Solo and Alto models noted, Solo runs on a single GPU and Alto on eight. Neither their size nor their training data was specified in the announcement reported by ICTjournal, which for now rules out any comparison with Apertus. Our section on AI tools and models follows these launches.

Is Apertus as good as commercial models?

That is not its goal, say its authors. "Apertus does not seek to compete with the frontier models of private companies, but to provide reliable and transparent AI," says Imanol Schlag in the CSCS announcement.

The published benchmarks place it on a par with open models of the same size. On general understanding tests, the Hugging Face model card for the first version gives Apertus 70B an average of 67.5%. Llama 3.1 70B scores 67.3% and Qwen2.5 72B 69.8%. These figures concern the model before fine-tuning, and the Apertus 1.5 announcement does not publish new ones.

Understanding dialect is, moreover, no longer a criterion that sets models apart. On 30 October 2026, the website Inside IT reported on a test by the company Intochat using 50 customer messages in Swiss German. The five commercial models tested captured 97.8 to 100% of the essential information. Above all, the same test found promises invented on behalf of the companies. Apertus was not included.

The model card points to the same limitation: Apertus can produce inaccurate or biased content, and should serve as an assistive tool, not a definitive source.

How can an SME or an individual try it?

Four access routes emerge from the official announcements and the Liip blog post, from the simplest to the most controlled:

  • the Public AI chat service, which offers the 8 billion version without registration and the 70 billion version with a free account, hosted under Swiss jurisdiction according to Liip;
  • Swisscom's Swiss AI platform, open to its business customers since the 2025 release, according to EPFL;
  • Infomaniak's API, which offers Apertus 1.5 with 70 billion parameters;
  • downloading from Hugging Face, to run the model offline on your own machines.

Local installation requires expertise: EPFL warns that it takes servers, cloud infrastructure or a dedicated interface. According to Liip, the 8 billion version runs on a recent Mac, and users have got it working, compressed, on an 8 GB laptop graphics card.

For an SME, the main benefit lies in control over data: a model installed on premises sends nothing to a third party. In return, the company alone has to check its answers.

What we do not know yet

  • CSCS does not publish any benchmark results for Apertus 1.5 against other models in its announcement of 24 July 2026.
  • The output filter meant to remove from answers the personal data covered by erasure requests is not yet offered on the Hugging Face model card for Apertus 70B.
  • MeditronFO's MED.USE clinical trials have not yet delivered results on patients.

Sources