Have you ever wondered how language models like ChatGPT actually work? Often this remains a secret because the big tech companies strictly guard their data and algorithms.
A new chapter was opened in Switzerland in September: With Apertus There is a language model that is built from the ground up. Open and transparent developed. This is a milestone for science and for all of us.
Why is Apertus so special?
The name Apertus comes from Latin and means ‘open’. And that perfectly describes the philosophy behind it. This project, led by the EPFL, the ETH Zurich and the CSCS (the Swiss Supercomputing Centre) is a counter-draft to the ‘black box’ models. Here is the entire development process – from the architecture to the training data to the source code – freely accessible and comprehensively documented.
This transparency is not only a nice gimmick, it is a game changer. It allows researchers and developers to accurately understand how it works, adapt the model for their own projects, and reproduce the results. Martin Jaggi, Professor at EPFL, calls it a ‘sample of a trustworthy AI model’.
A language model for global diversity
One of the biggest problems of many large language models is their strong focus on the English language. But Apertus goes another way. It was built on gigantic 15 trillion words Training that out More than 1,000 languages come from. 40 % These data are Non-English language! This is a real special feature.
As a result, Apertus also has access to languages that are otherwise often ignored in the AI world, such as Swiss German language and Romansh. This makes the model not only more versatile, but also more culturally relevant, especially in the multilingual Swiss context.
Open Source – an impetus for the future
The initiators do not simply see Apertus as a finished product. Thomas Schulthess, Director of the CSCS, describes it rather as a ‘initiator of innovation’. It should be a basic technology that forms the foundation for new applications such as chatbots, translators or digital learning tools.
The model is available in two different sizes: A variant with 8 billion parameters for individual users and one with 70 billion parameters for larger projects. Both versions are under one Permissive open source license It can even be used for commercial purposes.
Access for all: From professionals to enthusiasts
Do you want to experiment with Apertus yourself? Access is uncomplicated for professionals and experienced users:
- You can view the models on the platform Hugging face download. There you will also find all important documentation.
- As a strategic partner, we also Swisscom The model about your Sovereign Swiss AI Platform available for companies.
- For all those who want to participate in a global movement for open AI, the model is also available via the Public AI Inference Utility accessible.
For all those who want to plunge directly into the matter, the Swiss AI Weeks It has been an ideal opportunity since the beginning of September. Hackathons take place here, where developers can try out the model and give valuable feedback for further development.
Transparency and ethics as basic principles
Another important aspect in the development of Apertus was the adherence to the highest ethical standards. The developers have made sure that the model is all Swiss data protection and copyright laws and the transparency requirements of the EU AI Act fulfilled.
The training was based solely on publicly available data. Strict filters were used to remove personal data and unwanted content. Apertus shows that you can build a powerful AI model without compromising on fundamental values such as transparency, responsibility and the common good.
A collaborative project
Apertus is a community work. It is the result of collaboration between researchers, engineers and students from all over Switzerland. The project was supported by an investment of over 10 million GPU hours on the supercomputer Alps the CSCS It is also funded by the ETH Board. As a strategic partner, Swisscom has also made a significant contribution to the financing.
As Antoine Bosselut of the EPFL points out, the publication of Apertus ‘is not an end point, but the beginning of a journey’. The goal is to create an open, trustworthy and sovereign AI foundation for the global common good.
EN 
DE 
ES 
FR 
IT