The National Center for Artificial Intelligence (Cenia) in Chile unveiled Latam-GPT, an ambitious project to develop the first collaborative ChatGPT model focused on Latin America and the Caribbean.

The initiative responds to the need for a model that more accurately reflects regional culture, history, and context, overcoming the limitations of existing LLMs. 

In initial tests, ChatGPT provided descriptions of Latin American culture highlighting the richness of indigenous, African, and European heritage, music, dance, festivals such as Carnival and the Feria de las Flores, as well as literature and muralism.

These responses demonstrated how a globally trained model can capture general aspects but lacks a deep understanding of local diversity. 

Latam-GPT seeks to fill this gap through a collaborative approach, involving more than 30 institutions and 60 experts from different countries in the region. Rodrigo Durán, manager at Cenia Chile, stated, “The importance of this model lies in embedding the idiosyncrasy and worldview of Latin America into the most transformative technology of recent decades.” 

What is the project about?

The project aims to establish strategic partnerships with countries such as Mexico, Colombia, Ecuador, Argentina, Peru, Uruguay, Costa Rica, Spain, and the United States, creating a dataset that already reaches 50 billion parameters, comparable to OpenAI’s ChatGPT-3.5.

The data, exceeding 8 TB, comes from public and private virtual libraries and will be filtered, organized, and evaluated before final training. 

The high-performance infrastructure at the University of Tarapacá in Arica, Chile, will be key for training the model. The university acquired a supercomputer with 12 servers and 8 Nvidia H200 GPUs, featuring an efficient cooling system powered by 99% renewable energy. This setup allows the initial training to consume 135 kWh, producing only 0.96 tons of CO₂ with no additional water consumption. 

Omar Florez, a Peruvian expert based in Chile, leads Latam-GPT’s pre-training. The team is currently filtering information, running small-scale training tests, and aligning the model to develop “common sense” capabilities to provide accurate, context-aware responses. 

The UTA infrastructure investment reached US$4.5 million. Chile’s Minister of Science and Technology, Aisén Etcheverry, emphasized that “having a Latin American ChatGPT model allows us to advance toward more sustainable and representative artificial intelligence for our region.” 

Latam-GPT will use open sources, respect service terms, anonymize personal information, and comply with intellectual property laws. Additionally, the model aims to minimize environmental impact through renewable energy and efficient cooling systems. 

With this initiative, Latin America takes a significant step toward technological independence in artificial intelligence and representing its cultural diversity in the digital era.