Swiss Universities to Launch Open-Source AI Language Model

In a groundbreaking move that positions Switzerland at the forefront of ethical AI development, ETH Zurich and EPFL have announced their collaboration on the country's first public, open-source large language model. The initiative, scheduled for release in 2025, marks a significant departure from the proprietary AI models dominated by American and Chinese tech giants. This Swiss-led project emphasizes scientific transparency and ethical development, setting a new standard for responsible AI innovation in Europe.

The ambitious project, operating under the Swiss AI Initiative launched in December 2023, represents a remarkable collaboration within the Swiss academic ecosystem. More than 70 AI researchers from ETH Zurich and EPFL are working together, coordinated by the new Swiss National AI Institute. This collaborative approach, involving over 10 Swiss institutions, demonstrates Switzerland's commitment to pooling its intellectual resources for technological advancement. The project showcases the country's ability to compete with global tech giants while maintaining its tradition of scientific excellence and precision.

At the heart of this innovative project lies the Alps supercomputer, a state-of-the-art facility equipped with more than 10,000 NVIDIA Grace Hopper chips. This powerful infrastructure will support the training of the language model on an extensive dataset encompassing code, mathematical data, and text across multiple languages. The technical specifications reflect Switzerland's commitment to building robust, scalable AI systems that can compete with existing models like ChatGPT, Claude, Gemini, and Grok. The focus on high-performance computing infrastructure ensures the model's capability to handle complex computational tasks while maintaining Swiss standards of precision and reliability.

A distinguishing feature of the Swiss model is its impressive linguistic capabilities, supporting over 1,500 languages in its training dataset. With a balanced approach of 60% English and 40% non-English content, the model reflects Switzerland's multilingual heritage and commitment to global inclusivity. As highlighted by EPFL's Antoine Bosselut, this emphasis on multilingual capability has been a priority from the project's inception. This approach ensures the model's relevance and applicability across diverse global contexts, potentially making it a valuable tool for international organizations and multilingual communities worldwide.

Setting itself apart in a sector dominated by proprietary systems, the Swiss model champions transparency and ethical AI development. As emphasized by ETHZ AI Center's Imanol Schlag, the fully open nature of the model enables high-trust applications and advances research into AI risks and opportunities. The commitment to making source code, weights, and training data publicly available and reproducible supports adoption across various sectors while ensuring accountability. This approach aligns with European values and regulatory requirements, potentially influencing global standards for AI development and deployment. The project exemplifies Switzerland's tradition of promoting transparency and ethical considerations in technological advancement.