On February 10, 2026, Chilean President Gabriel Boric personally unveiled Latam-GPT: the first large language model developed inside Latin America. The project is led by Chile’s National Center for Artificial Intelligence (CENIA), directed by Álvaro Soto, and pulled together 15 countries, more than 200 collaborators, and 33 institutional alliances — universities, foundations, libraries, and government bodies from Argentina, Brazil, Colombia, Ecuador, Mexico, Peru, Uruguay, and beyond. Funding came from CENIA’s own resources and the Development Bank of Latin America (CAF).
The positioning matters more than the launch ceremony. This is not a consumer chatbot. It is open infrastructure: a 70-billion-parameter model built on Meta’s Llama 3.1, released under an open license, free for companies and public institutions to download and customize, and available through the official site at latamgpt.org.
A Model Grown From 2.6 Million Documents
The training data is the whole point. Latam-GPT was trained on more than 8 terabytes of regional material — 2.6 million documents spanning 20 Latin American countries plus Spain. It currently covers Spanish and Portuguese, with Indigenous Latin American languages on the roadmap. The backdrop is a long-standing data gap: Spanish and Portuguese together reportedly account for only around 2–3 percent of the training data behind existing large models, so outputs naturally tilt toward a US and Global North worldview.
Soto puts it bluntly: “No matter how powerful the large models are, they cannot cover all aspects relevant to our reality.” His example is cultural blindness — models default to figures like George Washington and US-centric framings, rather than the historical figures, idioms, and local context a Latin American user actually needs.
The Sovereignty Pitch: At the Table, Not on the Menu
Boric framed the project at the level of sovereignty, saying Latam-GPT positions the region as “an active and sovereign player in the economy of the future.” His most quoted line from the launch: “We’re at the table — we’re not on the menu.” Tech Policy Press also cited his framing of the project as “defending our identity and our right to exist.”
That argument resonates in 2026. With frontier capability and pricing concentrated in a handful of US and Chinese labs, regions have only two real levers: regulate, or build. Latam-GPT chose to build, and chose open source — any government, media outlet, or startup can download the weights and fine-tune them without depending on a specific vendor’s API.
The Reality of a $550,000 Budget
The ledger, though, is brutal. Total reported investment sits around $550,000 — roughly $300,000 from CENIA and $250,000 from CAF — which is more than three orders of magnitude below the multi-billion-dollar training budgets of US and Chinese labs. The first version was trained on AWS; future training runs are planned for a supercomputer at the University of Tarapacá in Arica, northern Chile, backed by a $10 million investment in the region’s first computing center capable of training large models domestically.
Use cases are already concrete. Soto points to hospital logistics as a candidate workload, and entrepreneur Roberto Musso’s company Digevo plans customer-service tooling that recognizes regional “slang, idioms, and even speech rate” — markets global models don’t serve well but that plainly exist.
Against the Global Open-Weights Backdrop
Skeptics raise two risks. First, OpenAI, Google, and Anthropic hold overwhelming resources and could cheaply fine-tune existing models on Spanish and Portuguese data, erasing Latam-GPT’s regional edge. Second, regional political instability threatens long-term continuity — the failed 2009 UNASUR fiber-optic project is the local precedent everyone brings up. Tech Policy Press’s verdict: Latam-GPT’s real value lies in capacity building and regional expertise, not in going head-to-head with global giants.
Set against the global open-weights landscape, that reading gets more interesting. The US and China keep shipping open-weight models on the performance track — MiniMax’s M2.5 open-weights coding model is a good example — while Latam-GPT runs the data-sovereignty track. For product teams the takeaway is practical: if you serve Spanish- or Portuguese-speaking markets, there is now a fully customizable base model with no API bill attached. Its value isn’t beating GPT at benchmarks; it’s that nobody can take it away from you.
Sources
- LatamGPT Navigates the Gap Between Regional Aspiration and Market Realities — Tech Policy Press
- Latam-GPT: a Latin American AI to combat US-centric bias — France 24
- Chile leads launch of first AI system developed in Latin America — Latin America Reports
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.