Swiss AI Initiative ships Apertus, a fully open multilingual foundation model

EPFL, ETH Zurich, and the national supercomputing center CSCS have released Apertus, a foundation model built under the Swiss AI Initiative that opens far more than its weights. The project publishes its training data, code, methods, and alignment principles, positioning itself as a reproducible, end-to-end transparent alternative to models that release weights alone. An accompanying Apertus Mini lineup of 16 small models showcases distillation and quantization approaches. The team claims parity with leading open models at the 8B and 70B parameter scales and trained the system on more than 1,000 languages from the start.

The more distinctive angle is regulatory and political rather than purely technical. Apertus is explicitly engineered to satisfy EU AI Act obligations: it honors data opt-outs, strips personally identifiable information, and is designed to limit memorization of training data. That compliance-first posture, combined with public European institutional backing and Swisscom as a strategic partner, frames the model as infrastructure for ‘sovereign AI’—capability that governments and organizations can audit and run without depending on US or Chinese providers.

For technical readers, the significance is less about benchmark wins and more about the openness contract. Most so-called open models keep their data pipelines and alignment recipes private, which makes independent auditing and reproduction impossible. By documenting the full stack, Apertus offers a base others can legally and verifiably build on, and a test of whether genuinely open, regulation-aligned models can stay competitive with their more closed counterparts.