The neutral Network

Why transparency could be the better AI strategy

On 2 September 2025, APERTUS was released – the country's first fully transparent language model. What may seem like naive Swiss good nature to some could prove to be a clever move – and change the way we work with machines.

OpenAI's models are anything but ‘open’. Google & Co. also guard their secrets like the apple of their eye. And even those who claim to be transparent are not really so. What ETH Zurich, EPFL and the CSCS supercomputing centre are doing with APERTUS is therefore almost a minor revolution in the AI market: with APERTUS, everything is open – training data, source code, model weights, even the intermediate steps of training.

This stands in stark contrast to the proprietary approaches of the major providers. And it raises a fundamental question: does transparency change the way we work with AI?

Digital commons vs. commercial products

APERTUS explicitly does not see itself as a competitor to ChatGPT, but rather as a platform for anyone who wants to use AI on their own terms. The Swiss scientists have deliberately chosen not to provide a finished product, but rather a technical foundation for further development by others.

This philosophy is more than just idealism, because Switzerland has a real structural problem: Silicon Valley and Beijing are competing for supremacy in AI, while Europe is in danger of being stuck in the digital second division. Those who cannot keep up with the arms race for the most powerful AI must find other ways to remain relevant – Apertus could be one of them.

Collaboration instead of monopolisation

The technical data is impressive: 4,096 GH200 GPUs, 15 trillion training tokens, over 1,000 languages. The new xIELU activation method and the AdEMAMix optimiser are proprietary developments of the Swiss teams.

However, from our perspective, the decisive difference is another: 40 per cent of APERTUS' training data is non-English. This is a conscious rejection of the Anglophone dominance of the AI world – and a direct attack on one of the biggest problems in everyday work with AI.

Language as the key

This is where APERTUS becomes interesting for practitioners: the model speaks Swiss German and Romansh. For Swiss people, this means no more detours via High German or English, and fewer cultural translation losses.

This is more than just a technical gimmick. Anyone who works with AI on a daily basis is familiar with the problem: the quality of collaboration depends crucially on how naturally communication works. If I first have to translate what I mean four times to a system and receive answers that seem poorly cobbled together, friction losses arise. Trust in the collaboration is gone.

Three practical advantages

For anyone who works with AI professionally, APERTUS offers three concrete improvements:

• Better communication: Natural language leads to more productive collaboration. No translation loops, fewer misunderstandings.

• More trust: When I know what a model has been trained on, I can better assess its responses. This is particularly important when making critical decisions.

• Genuine customisation: APERTUS can be modified, expanded and adapted to specific needs. This opens the door to tailor-made solutions that would be impossible with black box systems.
  

A statement with a signal effect

APERTUS is more than just another language model. It is a statement about what responsible AI development should look like: collaborative instead of monopolistic, transparent instead of secretive, culturally sensitive instead of hegemonic. The power of language models is too great for us to leave them in the hands of a few ambitious businesspeople.

The Swiss have proven that there is another way. At a time when AI is often romanticised as magic, they are taking the opposite approach and allowing us to look into the heart of the machine. Some magicians on the market may not like this, but for most of us it is a gift – and an encouraging signal: because the best technology is not the most mysterious, but the one we can trust the most.

The coming months will show whether transparency works as a business model. And whether the Swiss will benefit from the open approach of their research. But one thing is already clear: APERTUS is changing the rules of the AI game, and that could benefit us all.

---

APERTUS is available via Hugging Face, the Swisscom network and the Public AI network. The Apache 2.0 licence allows commercial use without any hidden pitfalls. And I continue to use em dashes for stylistic reasons, whether the algorithm likes it or not.