The Swiss AI Initiative, a collaboration between EPFL, ETH Zurich, and CSCS, has developed the 'Apertus' foundation model based on open weights, data, and science. Designed to meet EU AI Act requirements, the model is competitive with top open models at an equivalent scale of 8B and 70B parameters.
The Swiss AI Initiative, a collaboration between EPFL, ETH Zurich, and CSCS, has publicly released the open foundation model 'Apertus'. The model is built on principles of open weights, open data, and open science, with all training data, code, weights, methods, and alignment principles documented and reproducible.
Apertus is designed to meet the requirements of the EU AI Act, respecting opt-outs, removing PII, and preventing model memorization. It is competitive with top open models at an equivalent scale of 8B and 70B parameters and is multilingual from day one, trained on 1000+ languages. Swisscom is a Strategic Partner of the initiative.
This release is a significant milestone for European efforts in sovereign and transparent AI development. By applying open-source principles to a foundation model, it establishes a global foundation for innovation while ensuring regulatory compliance.
Some users criticize the previous version's performance and its copyright compliance claims, while others emphasize that its full openness (weights, data, training details) constitutes a significant scientific contribution. Active comparisons are made with other open models like Nemotron and OLMo, and one comment argues that 'sovereign AI' is about using multiple models collaboratively rather than relying on a single one.