Latam-GPT is an innovative large language model being crafted specifically for the Latin American region. Spearheaded by the nonprofit Chilean National Center for Artificial Intelligence (CENIA), this initiative aims to promote technological autonomy in the area by developing an open-source AI model designed for the languages and contexts of Latin America.
“This venture cannot be carried out by a single individual or nation within Latin America; it’s a challenge that necessitates collaborative involvement,” states Álvaro Soto, director of CENIA, during an interview with WIRED en Español. “Latam-GPT is focused on establishing a free, open, and cooperative AI model. We’ve been engaged for two years with a grassroots methodology, involving citizens from various nations who are keen to participate. Recently, governments have also expressed interest and begun to engage in the project.”
The project stands out for its collaborative spirit. “Our aim isn’t to rival OpenAI, DeepSeek, or Google. We intend to develop a model tailored to the needs of Latin America and the Caribbean, addressing cultural requirements and challenges, such as grasping dialects, the region’s historical context, and distinct cultural characteristics,” Soto elaborates.
With 33 strategic partnerships across Latin America and the Caribbean, the initiative has compiled a data corpus that exceeds eight terabytes of text, equivalent to millions of books. This extensive data has enabled the development of a language model featuring 50 billion parameters, making it on par with GPT-3.5 and capable of sophisticated tasks such as reasoning, translation, and making associations.
Latam-GPT is trained on a regional database comprising information from 20 Latin American countries and Spain, amounting to 2,645,500 documents. The distribution of this data reveals a concentration in the region’s larger nations, with Brazil leading with 685,000 documents, followed by Mexico with 385,000, Spain with 325,000, Colombia with 220,000, and Argentina with 210,000 documents. These numbers reflect the size of these markets, their digital advancement, and the availability of structured content.
“We will initially release a language model. Its performance on general tasks is anticipated to be comparable to leading commercial models, but it will excel in topics pertinent to Latin America. The intention is for its understanding of region-specific topics to be much more profound,” Soto articulates.
The initial model will serve as a cornerstone for advanced future technologies, including those related to image and video, and for scaling to more extensive models. “As an open initiative, we encourage other organizations to make use of it. A group in Colombia may tailor it for educational purposes or one in Brazil for healthcare. The goal is to empower various institutions to develop specialized models for sectors like agriculture, culture, and beyond,” states the CENIA director.
