Quick Answer:
Latam-GPT is the first large language model built for Latin America. According to Tech Policy Press, it was launched on February 10, 2026 by CENIA as open-source infrastructure rather than a direct-use chatbot, was built on Meta's Llama 3.1 as a 70-billion-parameter model, and was trained on more than 8 terabytes of Spanish and Portuguese data.
Key Takeaways:
- Launch: According to Tech Policy Press, Latam-GPT was launched on February 10, 2026 as open-source infrastructure — a foundation for developers, not yet a direct-use chatbot.
- Regional leadership: The project is led by CENIA and involves 15 countries, more than 200 collaborators, and 33 institutional alliances.
- Technical scale: It is a 70-billion-parameter model, built on Meta's Llama 3.1, trained on more than 8 terabytes of data (2.6 million documents).
- The problem it solves: Spanish and Portuguese material makes up only 2 to 3 percent of existing AI training material, according to Tech Policy Press.
- Investment: The University of Tarapaca invested $10 million in a supercomputing center; CENIA allocated $300,000 and CAF committed $250,000.
For years, the artificial intelligence that Spanish-speaking businesses relied on was built in another language, for another market. The dominant models learned from English, from a US context, and from the nuances of a consumer who does not always resemble the customer walking into your shop in Houston, Cypress, Monterrey, or Bogota. That gap is exactly what a regional project set out to close.
According to Tech Policy Press, Chile launched Latam-GPT, described as the first large language model built specifically for Latin America. For any business owner who depends on AI understanding their customers — in their language, in their context — this is a development worth understanding.
What Latam-GPT Is and Who Built It
According to Tech Policy Press, Latam-GPT was launched on February 10, 2026 as open-source infrastructure — not a direct-use chatbot, but a foundation that developers and institutions can build on. The project is led by CENIA, Chile's National Center for Artificial Intelligence, which is directed by Alvaro Soto. That distinction is why the region-built open model is still maturing toward a fully usable release: the open foundation is out, and the work of turning it into products people use every day is the road ahead.
This is not a single-country effort. Tech Policy Press reports that the project involves 15 countries, more than 200 collaborators, and 33 institutional alliances across the region. It is a collaborative undertaking at a scale that is uncommon for AI infrastructure in Latin America.
On the technical side, according to Tech Policy Press, Latam-GPT is a 70-billion-parameter model built on Meta's Llama 3.1. Rather than starting from scratch, the team took an open base model and adapted it with regional data — a practical route for a project pursuing technological self-determination without the cost of training a model from the ground up.
The number that explains it all: According to Tech Policy Press, Spanish and Portuguese material makes up only 2 to 3 percent of existing AI training material. That scarcity is why global models so often stumble over the region's slang, cultural references, and market context.
The Data Scale Behind the Model
Building a model that genuinely understands the region demands data from the region. According to Tech Policy Press, the development of Latam-GPT processed more than 8 terabytes of data, equivalent to 2.6 million documents, with training data drawn from 20 Latin American countries plus Spain.
That volume matters for a concrete reason: a model trained mostly on English text learns the world through a US lens. When a customer in Latin America phrases a question using local idioms, regional product names, or culturally specific references, a global model can misread it. A model trained on data drawn from 20 Latin American countries plus Spain has, at least in theory, a better chance of getting it right.
The decision to build on an open base model is worth pausing on. According to Tech Policy Press, Latam-GPT was built on Meta's Llama 3.1 rather than trained from scratch. That choice tells you something about the strategy: instead of competing dollar-for-dollar with the compute budgets of the largest commercial labs, the project layered regional data onto an existing open foundation. According to Tech Policy Press, global AI remains dominated by OpenAI, Google and Anthropic, with resources several orders of magnitude larger than this regional effort. Adapting an open model is how a region working with far smaller budgets can still produce something tailored to its own languages and context.
AI Sovereignty: the Underlying Argument
The framing the project's leaders gave it goes beyond technology. Alvaro Soto, the director of CENIA, articulated the technical limitation that justifies a regional model. According to Tech Policy Press, Soto said: "No matter how powerful the large models are, they cannot cover all aspects relevant to our reality."
The argument is one of technological self-determination: a region whose languages and context are underrepresented in the dominant models building its own foundation rather than depending entirely on systems made elsewhere. According to Tech Policy Press, Spanish and Portuguese make up only 2 to 3 percent of the training material in existing AI models — the gap a region-built model is meant to close.
Tech Policy Press also places the project in historical context, pointing to the UNASUR fiber-optic network — a regional technology effort announced in 2009 that collapsed in 2019 amid political disagreements and funding shortfalls — as a cautionary precedent. The open question is whether an AI initiative can sustain the cross-border cooperation that the earlier infrastructure project could not.
The investment behind the project, according to Tech Policy Press:
- The University of Tarapaca in Arica, Chile invested $10 million in a supercomputing center.
- CENIA allocated $300,000 to the project.
- The Development Bank of Latin America and the Caribbean (CAF) committed $250,000.
Regional Aspiration Versus Market Reality
The enthusiasm for a region-built model is real, but so are the challenges. Tech Policy Press frames the project squarely around the gap between regional aspiration and market reality: funding, access to compute, and adoption.
An open model, on its own, does not guarantee users. For Latam-GPT to change how businesses operate, developers have to build on it, companies have to integrate it, and the dominant commercial models — already entrenched — have to make room. That is the road Tech Policy Press indicates the project still has to travel.
For business owners, the lesson is not to wait and see which model wins. The lesson is that AI, in all its forms, is becoming the layer through which customers discover businesses. And that raises a more urgent question than which model to pick: when a customer asks AI for a service like yours, does your business appear in the answer?
What This Means for Your Business
Latam-GPT is a signal of a larger trend: AI is learning to understand regional Spanish and Portuguese better and better. As that happens, more consumers — both across Latin America and within the Hispanic community in the United States — will look for businesses by asking an AI assistant directly, in their own language. If the AI does not know your business, it cannot recommend it.
This is where Answer Engine Optimization (AEO), the MerchandisePROS service built precisely for this moment, comes in. AEO structures your website, your business data (schema), your FAQ content, and your directory citations so that models like ChatGPT, Perplexity, Google AI Overviews — and the regional models now arriving — can find, understand, and cite your business with confidence. The problem it solves is concrete: being invisible to the AI layer that more and more customers pass through before they decide.
The first step is knowing where you stand today. Our free audit checks your digital presence across more than 20 factors — including your visibility to AI — and delivers a 0-to-100 score with a clear action plan, in under 60 seconds.
Frequently Asked Questions
What is Latam-GPT?
Latam-GPT is the first large language model built specifically for Latin America. According to Tech Policy Press, it was launched on February 10, 2026 as open-source infrastructure (not a direct-use chatbot), is led by CENIA, and was built on Meta's Llama 3.1 as a 70-billion-parameter model trained on Spanish and Portuguese material.
Who is behind Latam-GPT?
According to Tech Policy Press, the project is led by CENIA, Chile's National Center for Artificial Intelligence, directed by Alvaro Soto, and involves 15 countries, more than 200 collaborators, and 33 institutional alliances across the region.
Why does an AI model trained in Spanish and Portuguese matter?
According to Tech Policy Press, Spanish and Portuguese material makes up only 2 to 3 percent of existing AI training material. A model trained on regional data promises to capture the context, language, and market nuance that US-centric models often miss.
How large is Latam-GPT's training dataset?
According to Tech Policy Press, development processed more than 8 terabytes of data, equivalent to 2.6 million documents, with training data drawn from 20 Latin American countries plus Spain.
What does Latam-GPT mean for my business?
As more AI models understand regional Spanish and Portuguese, more customers will look for businesses by asking AI in their own language. Answer Engine Optimization (AEO) structures your site and data so those models cite your business. The free MerchandisePROS audit checks how visible you are to AI.
"It does not matter which AI model wins. What matters is whether the AI knows your business well enough to recommend it when a customer asks in their language."
- Diego Medina F, Founder of MerchandisePROS