Chile Launches Regional LLM to Counter US Bias
Chile, in partnership with AWS and regional institutions, has launched Latam-GPT, the first open-source large language model trained specifically on data from Latin America and the Caribbean. The initiative is a direct effort to counter the US-centric bias found in many global AI models and establishes a precedent for regional sovereignty in foundational AI infrastructure.
- The project is a major collaborative effort involving over 60 institutions and nearly 200 specialists from more than 15 countries in Latin America and the Caribbean. Key partners include Chile's National Center for Artificial Intelligence (CENIA), the Development Bank of Latin America and the Caribbean (CAF), and Amazon Web Services (AWS). - Latam-GPT is being built using Meta's open-source Llama 3.1 architecture with 70 billion parameters. Its first version, scheduled for release in September 2025, is being trained on 70 billion words in Spanish and Portuguese. - The initial development was notably cost-efficient, funded with just $550,000 primarily from CENIA and the Development Bank of Latin America (CAF). Future training will leverage a supercomputer at the University of Tarapacá in northern Chile. - A primary goal is to incorporate and preserve Indigenous languages, which are often underrepresented in major AI models. The initial phase will focus on including Rapa Nui and Mapudungun. - The model is trained on over eight terabytes of regional and synthetic data, equivalent to millions of books, to better capture local slang, idioms, and cultural nuances. This data was sourced from universities, libraries, government entities, and civil society organizations across the continent. - The initiative is led by Álvaro Soto, the director of Chile's National Center for Artificial Intelligence (CENIA), and has received public endorsement from Chilean President Gabriel Boric. - Beyond countering bias, the project is framed as a move toward "digital sovereignty," allowing the region to build and shape its own foundational AI models rather than only consuming technology developed elsewhere. - Latam-GPT will be available free of charge for companies, universities, and public institutions to build applications tailored to Latin American needs, such as in healthcare and public services.
