Fact and fiction about the Swiss AI model Apertus 

Since its launch in early September, much has been said and written about the Swiss AI model Apertus. What is true and what is false? Keystone / Gaetan Bally

The Swiss federal institutes of technology recently launched Apertus, a large-scale artificial intelligence (AI) model. Reaction has been mixed. We look at what’s behind the most common claims about the new technology.

After weeks of anticipation, ETH Zurich and EPFL recently unveiled Apertus, a national large-scale language model (LLM). Developers describe it as “fully open” – meaning that all parts of the model are openly available to the public – and “a milestone in generative AI for transparency and diversity”. Various claims are circulating about what it can or can’t do. What is true and what is false?

Swissinfo tested Apertus and asked the developers behind it, as well as a few other experts in the AI field, to separate fact from fiction and shed light on its main strengths and weaknesses.

>> Why the name Apertus reinforces societal prejudices, according to Claudia Vaccarone:

Apertus is Switzerland’s ChatGPT: False

Apertus is not designed for personal use like OpenAI’s ChatGPT. It’s a foundation AI model developed to be adapted to different applications and services, especially in business and research. For example, it could be used to power e-commerce or extract information in multiple languages for medicine.  

“The general public is not our main target audience,” says Imanol Schlag, a researcher at ETHZ who directed the technical development of Apertus.

For this reason, Apertus was not released with many user-facing features or a dedicated mobile app.

“Our intention was never to develop the Swiss answer to ChatGPT,” explains Maria Grazia Giuffreda, associate director of the Swiss Centre for Scientific Computing in Lugano, where the supercomputer that runs Apertus is based.

To give the public the chance to try out the new AI, the NGO Public AI has made some chat functions available on its websiteExternal link.

>> Read all about what’s new in the field of AI in Switzerland:

Apertus can’t compete with the most widely used LLMs: True

Apertus is the most powerful fully open LLM to our knowledge, with 70 billion parameters (a value indicating learning capacity) in its largest version. But it doesn’t match proprietary models such as GPT-4, Gemini or Claude, which are trained on far larger datasets and backed by massive computing power. By comparison, GPT-3 has 175 billion parameters (details are not public for later versions).

“Comparing Apertus to big American companies’ AI models is like comparing a small farmer in Valais to a massive beef producer,” says El Mahdi El Mhamdi, professor at the École Polytechnique of Paris.

Smaller models, however, can be more efficient, more accessible to small and medium-sized enterprises (SMEs) and less energy-intensive. “There is a growing realisation that data quality is more important than quantity,” adds Mete Ismayilzada, researcher at EPFL and at the University of Italian-speaking Switzerland.

Some public intellectuals remain sceptical. Bruno Frey, professor emeritus at the University of Zurich, said the Swiss LLM gave him an incorrect scientific source. “I didn’t find it very convincing,” he said. Mathematician Xavier Comtesse wrote on LinkedInExternal link that the model produces many results that look accurate but are actually wrong (hallucinations).

Schlag defends Apertus’s performance, noting that its smaller version (eight billion parameters) already outperforms comparable models from Mistral, Meta, Alibaba – and every other public institution.

>> This video helps you navigate through the technical terms concerning AI:

Apertus is ethical and transparent compared with other AI models: True

Apertus is the first large-scale AI model designed and developed to comply with the key requirements of the European AI Act, which took effect in 2024. These include transparency, data traceability and respect for intellectual property and privacy. The model architecture, the weights (numerical values that determine how the neural networks work) and training instructions are public. The datasets used for training come from public and legal sources and do not include copyrighted works or sites that have opted out of being included in the training datasets.

By contrast, tech giants have long exploited the data of billions of users without consent, including copyrighted ones. This data is often compiled using hidden underpaid human labourExternal link in developing countries, notes El Mahdi El Mhamdi. “We do not yet realise the level of abuse that is normalised in the supply chain of modern mainstream AI,” he says.

That’s why Apertus is particularly attractive for companies, research institutions and public bodies that want to develop AI applications ethically and in compliance with legal requirements. “We are an example that you can train generative AI responsibly and without stealing other people’s intellectual property,” Schlag says.

Apertus speaks more than 1,800 languages: Misleading

Developers have made a great effort to train Apertus on a great variety of languages – more than 1,800, according to them, including minority languages and dialects often overlooked by popular LLMs such as Romansh and Swiss German. This is significant, as most AI models focus primarily on widely spoken languages.

But understanding a language is one thing, speaking it correctly another. Especially in less common languages, Apertus can make noticeable mistakes.

In our test, for example, Apertus sometimes produces awkward or incorrect sentences in Italian. In Romansh, it gave a wrong translation for the word “grandfather”.

Apertus’s head of development Schlag admits that the model’s conversational abilities need to be improved. However, he claims it already outperforms other language models in certain tasks, for instance in German-Romansh translation, according to the latest technical reportExternal link.

This doesn’t impress everyone. “What’s the point of a model that speaks Romansh if its average performance is far below that of comparable models – not to mention the top performers?” asks Aldo Podestà, CEO of Swiss AI start-up Giotto.ai, who supports the Apertus project but also recognises its limitations.

Mary-Anne Hartley, professor and director of the EPFL’s laboratory for intelligent global health and humanitarian response technologies, disagrees with this view. “People who speak languages that are not widely spoken deserve to be represented in technology. This is what Apertus is trying to do,” she says. For developers, these early flaws are the temporary price of a more inclusive and ethical approach.

Apertus is only meant for Switzerland: False

Although it was developed in Switzerland by Swiss institutions, it would be wrong to see Apertus as useful only for applications in Switzerland. Most of the training datasets come from international sources. The exceptions include data in Swiss German and Romansh, and the integration of a “Swiss values charter” that lays out principles such as neutrality and linguistic diversity to which the AI must align. “Apart from these aspects, our model has nothing else specifically Swiss,” says Imanol Schlag.

The Apertus team hopes that other countries will show interest in this project and help advance its development by providing infrastructure, talent and resources. “Our ambition is to continue to develop Apertus at a European if not global level,” says  Giuffreda of the Lugano computing centre.

Apertus can’t update itself in real time because it’s not connected to the internet: Misleading

No large language model can update itself in real time. They all remain static once trained, even when embedded in products that can access the internet, such as ChatGPT. The only way to insert changes and corrections in an AI model is to re-train it. But this is a very expensive process that only companies with large resources can afford to do frequently. “That is the main limitation of this technology,” Schlag explains.

For the next training round, Apertus will draw on CHF 20 million ($25 million) in federal funding and the Swiss Alps supercomputer, powered entirely by hydropower, thus saving resources. But in the long run, new funding sources will be needed. “I would like to see more investment in this technology, which is so crucial for our digital sovereignty,” Schlag says.

Edited by Gabe Bullard/ts