AI models are transforming how we create, share and preserve culture. At the same time, there is a homogeneity issue when models are trained on primarily English data and present biased information about regional realities. When developed with inclusivity in mind, large language models (LLMs) have a real potential to expand access, preserve heritage, strengthen diverse representation and inspire new forms of creativity. This is exactly the idea behind Latam-GPT.
Latam-GPT represents Chile's first open source LLM trained by the cultures of Latin America. The AI model was first announced in February 2026 at the Artificial Intelligence Action Summit in Paris, but its development started around 2023 when a research team at the National Center of Artificial Intelligence (CENIA) in Chile began exploring ways to develop a regional-specific LLM built on existing open source AI models.
Right from the beginning, the team at CENIA wanted to “learn and build a platform that was open source and that paid attention specifically to the cultural language-based difference, but also the cultural history and points of view within the Latin American region”, explains Gabriela Arriagada Bruneau, Associate Researcher at CENIA and assistant professor in AI and data ethics at the Pontifical Catholic University of Chile.
Addressing the representation gap
Latam-GPT does not intend to compete with ChatGPT, Google Gemini and similar consumer tools. Instead, the model aims to offer a more accurate and efficient foundational infrastructure for the further development of regional AI applications. The vision behind the model became clear when the team at CENIA examined the data behind popular LLMs.
“We wanted to see what level of representation existed, and roughly under two percent of all the training data and training models we considered contained actual Spanish or Portuguese data”, Arriagada Bruneau explains.
The team also identified issues that went beyond simply increasing Spanish and Portuguese content. Another key issue was linked to how AI models represent historical and cultural differences within Latin America.
“It was not just about processing language, but also more profound questions,” she says. “How should a model refer to historical differences? We know that how you train the model and how you put any type of filtering or benchmarks is going to have an impact on how the output is going to represent those things.”
To efficiently address these issues, the project relies heavily on collaboration. CENIA works with partners from more than 20 Latin American countries, including academic and governmental institutions, non-profit organisations and local communities.
Knowledge testing through community engagement in benchmarks
Latam-GPT’s collaborative approach is also shaping how the model is evaluated. By developing its own benchmarks, the team is able to test the extent to which language models accurately understand regional knowledge.
“We’ve built benchmarks, and I think this has been a very beautiful way of learning,” says Arriagada Bruneau. “It has come up as a very nice community embrace of how we think about language models.”
One of these is Choclo, an open source benchmark centered around Latin American cultural knowledge. Choclo intends to capture the richness, diversity and specificity of Latin American cultural knowledge. It contains over 100,000 rows of questions covering topics such as geography, fauna, flora, traditions, gastronomy and public figures. This approach is not just about general knowledge that LLMs normally acquire from the internet, but also about engaging people to provide this knowledge.
Besides Choclo, the team also developed the open source benchmark Trueque, consisting of 500 curated questions to test factual knowledge and cultural appropriateness. To further train Latam-GPT, the experimental interactive chat platform Copuchat was created to train the model.
Openness, data governance and AI ethics
As explained by Arriagada Bruneau, openness is central to the project, which includes both open source and open weights. The team is also focusing on questions around data governance and AI ethics, including transparency, privacy, bias and linguistic diversity.
“You have what I would call ‘traditional ethical principles’ – transparency, responsibility, fairness, accounting for bias” she says. “But we also considered, for example, criteria for linguistic diversity”.
The team at CENIA have also developed a transparency sheet detailing the reasoning behind the project’s data interventions and privacy levels that are being respected, in collaboration with the GobLab UAI team. Because of a limited AI regulation in Chile, much of the project’s responsible AI framework is based on the team’s own commitment to openness and AI ethics.
Second phase: strengthening Latam-GPT and further regional collaboration
As the project enters its second phase, the focus is on strengthening the reasoning and problem-solving capabilities of Latam-GPT as well as expanding contextual training. CENIA is also working with anthropologists, sociologists and consulting representatives from various indigenous communities to explore the possibilities to include indigenous languages and knowledge.
To further strengthen regional collaboration, the team is building a strategic alliance with Brazil. Such collaborations allow the project to expand its technical capabilities, but they also help challenge the assumption that AI development has to follow a predetermined path.
This may ultimately be one of Latam-GPT’s most important contributions. The project moves away from simply adding more Latin American data to AI models and instead focuses on exploring what it actually means to develop a model within the region: “there are different ways in which we can make technology our own to serve the needs and the contexts of different people”.
