Apertus – Open Foundation Model for Sovereign AI (apertvs.ai)

Apertus has introduced an open foundation model designed for sovereignty in AI, featuring a suite of 16 small language models that showcase advanced distillation and quantization techniques. The release emphasizes transparency, with all training data, coding methodologies, model weights, and alignment principles made fully documented and reproducible. Notably, Apertus adheres to the EU AI Act, ensuring compliance by implementing measures for personal data protection, opt-out capabilities, and preventing memorization of sensitive information. This makes it a foundational tool for responsible AI development. Significantly, Apertus competes with leading open models like the 8B and 70B parameter systems, demonstrating multilingual capabilities from the outset, having been trained on over 1,000 languages. The Swiss AI Initiative, in partnership with Swisscom, endorses this effort, and its accompanying technical report has gained acceptance at a prestigious AI and NLP conference. This release not only enhances the accessibility of robust AI models but also sets a precedent for responsible AI practices globally, fostering innovation while prioritizing ethical considerations in AI deployments.