Latam-GPT: How Chile Built Latin America’s First Sovereign AI Model

A Continent Builds Its Own AI

In the global AI race, the narrative has been dominated by a handful of players: OpenAI and Anthropic in the United States, Google DeepMind between London and Mountain View, Mistral in Paris, and a cluster of ambitious labs in China. The rest of the world — the vast majority of humanity — has been cast as consumers of AI built elsewhere, trained on data that reflects someone else’s language, culture, and priorities.

Chile just challenged that narrative. On February 10, 2026, the National Center for Artificial Intelligence (CENIA) unveiled Latam-GPT in the presence of Chilean President Gabriel Boric. Built on Meta’s Llama 3.1 architecture and trained on more than 300 billion tokens of Latin American data in Spanish and Portuguese, Latam-GPT is the first open foundation model created entirely within Latin America — from data collection and pre-training to post-training. Backed by a consortium of more than 60 institutions and nearly 200 specialists across eight core Latin American countries, the project was presented jointly by CENIA, CAF (the Development Bank of Latin America and the Caribbean), the Government of Chile, AWS, and the Data Observatory.

The most remarkable number in the Latam-GPT story may not be the model’s scale but its budget: $550,000 — funded primarily by CENIA and CAF. In a field where frontier model training runs routinely cost hundreds of millions of dollars, the project demonstrates what focused, collaborative effort can achieve even with modest resources.

The project is not just a technical achievement. It is a political statement about who gets to shape the AI systems that increasingly mediate access to information, services, and opportunity. In a world where the dominant AI models are trained primarily on English-language data — Spanish accounts for roughly 4% of typical LLM training data, Portuguese just 2% — Latam-GPT is an assertion that Latin America’s roughly 660 million people deserve AI that understands their languages, their contexts, and their needs.

The Technical Architecture

Latam-GPT was not trained from scratch — an undertaking that would have required billions of dollars in compute, far beyond any Latin American institution’s budget. Instead, the project took a pragmatic approach: it started with Meta’s Llama 3.1 as a foundation and performed an extensive continued pre-training campaign focused on Latin American data.

The training dataset, curated over more than two years by teams across the consortium, includes government documents from Latin American countries, academic papers from regional universities, court decisions, library records, school textbooks, news articles from major Latin American outlets, literary works in the public domain, legal texts, and curated web content in Spanish and Portuguese. The total dataset exceeds eight terabytes — comprising more than 300 billion plain-text tokens, equivalent to around 230 billion words. Modest by the standards of frontier model training but substantial for a regional effort, and critically, highly focused on the specific linguistic and cultural contexts that global models handle poorly.

The continued pre-training process adapted Llama’s existing knowledge to the specificities of Latin American Spanish and Portuguese. Latin American Spanish differs substantially from European Spanish in vocabulary, idiom, and register. Brazilian Portuguese and European Portuguese diverge even more sharply. A model trained primarily on European or generic Spanish text will misunderstand regional expressions, mishandle country-specific terminology, and produce outputs that feel foreign to Latin American users. Latam-GPT was specifically designed to close this gap. The training data also includes indigenous languages — Nahuatl, Quechua, and Mapudungun — as well as Caribbean dialect variants, though full support for these languages is planned for future versions.

Post-training, the model underwent instruction tuning and alignment using feedback from native speakers across multiple countries. This process ensured that the model not only understood Latin American text but could generate responses that felt natural and culturally appropriate to users in Mexico, Colombia, Brazil, Argentina, Chile, Peru, and other countries represented in the consortium.

The first version was trained on AWS cloud infrastructure, with a $4.5 million supercomputer planned for installation at the University of Tarapaca in northern Chile during the first half of 2026 to support future training runs. The model was released as an open model, with initial tooling available on Hugging Face. Critically, Latam-GPT is positioned not as a consumer chatbot but as foundational infrastructure — designed for text-intensive workflows common across public administration and services, including document drafting, summarization, translation, knowledge retrieval, and citizen support.

Why Sovereign AI Matters

The concept of “sovereign AI” — the development of AI systems that reflect and serve a specific nation or region’s interests — has gained tremendous momentum worldwide. France has championed Mistral as a European answer to American AI dominance. The Gulf states are investing heavily, with Abu Dhabi’s Technology Innovation Institute releasing Falcon 3 and Falcon-H1 Arabic — now the leading Arabic AI model — using a novel hybrid Mamba-Transformer architecture. India has launched Bhashini, a government platform supporting AI in 22 Indian languages that was recently migrated to domestic cloud infrastructure. Japan, South Korea, and Singapore have each announced national AI strategies with significant public funding.

The motivation is not mere technological nationalism. It reflects a genuine and well-founded concern about what happens when the AI systems that a society depends on are built elsewhere, by people with different values, different priorities, and different cultural contexts.

Language is the most obvious dimension. Despite progress in multilingual capabilities, frontier AI models remain substantially better in English than in any other language. They understand English idiom, humor, cultural references, and technical terminology with a depth that they cannot match in Spanish, Portuguese, Arabic, or Hindi. For the billions of people who do not speak English as a primary language, this means that the AI revolution delivers a degraded product — and the degradation is greatest precisely in the areas where cultural and linguistic specificity matter most.

But the concern goes beyond language. AI models encode values and assumptions in their training data and alignment processes. A model trained primarily on American and European data will reflect American and European perspectives on topics ranging from governance to economics to social norms. It may misunderstand or misrepresent local contexts, recommend inappropriate solutions to local problems, or simply lack knowledge about issues that are critical to non-Western societies.

Latam-GPT addresses these concerns directly. By training on Latin American data, with feedback from Latin American users, under the direction of Latin American researchers, the model is designed to serve the region on its own terms. It is not a translation layer on top of an American model — it is a model that has been adapted from the ground up to think in Latin American Spanish and Portuguese.

The Consortium Model

One of Latam-GPT’s most significant innovations is organizational rather than technical. Building a competitive AI model requires enormous resources — data, compute, talent, and funding — that no single Latin American institution could provide alone. The solution was a continental consortium that pooled resources across national boundaries.

The consortium includes universities, government agencies, and research institutes from Chile, Brazil, Mexico, Colombia, Argentina, Peru, Ecuador, and Uruguay as core data-contributing nations, with the broader network extending to up to 15 Latin American and Caribbean countries. Each partner contributed data, expertise, and in some cases compute resources. The coordination was managed by CENIA in Santiago, which served as the technical hub and primary training facility, with AWS providing cloud infrastructure for the initial training.

This model has advantages and challenges. The advantages are clear: resource pooling makes possible what no single institution could achieve. The diversity of contributing organizations ensures that the training data and evaluation criteria reflect the full breadth of Latin American cultures rather than a single country’s perspective. Brazil contributed Portuguese language expertise and Amazon region data. Mexico provided indigenous language specialists and educational content. Colombia contributed biodiversity and agricultural knowledge systems. The open release model means that the benefits flow to the entire region rather than accruing to a single commercial entity.

The challenges are equally real. Coordinating across more than 60 institutions in multiple countries requires navigating different regulatory frameworks, institutional cultures, and political dynamics. Data sharing across national boundaries raises privacy and sovereignty concerns. And the funding model — $550,000 from CENIA and CAF, supplemented by AWS cloud credits and institutional contributions — creates uncertainty about the project’s long-term sustainability. Frontier AI development is not a one-time investment but an ongoing commitment that requires continuous compute, data curation, and model updates. The planned $4.5 million supercomputer at the University of Tarapaca represents an important step toward infrastructure independence, but the gap between a $5 million investment and the billions being spent by US and Chinese AI labs remains vast.

Lessons for Africa and the Arab World

The Latam-GPT project carries profound implications for other regions that find themselves on the consuming end of the AI divide. Africa, the Arab world, and Southeast Asia all face similar challenges: linguistically diverse populations poorly served by English-centric AI models, limited domestic compute infrastructure, and growing dependence on AI systems built in Silicon Valley or Beijing.

The consortium model offers a template. No single African nation has the resources to build a frontier AI model, but a continental or sub-regional effort — pooling data from multiple countries, leveraging diaspora talent, and coordinating through existing institutions like the African Union or the Arab League — could potentially achieve what Latam-GPT has demonstrated.

For the Arab world specifically, the parallels are striking. Modern Standard Arabic is reasonably well represented in frontier model training data, but dialectal Arabic — the language that people actually speak — is vastly underrepresented. The UAE’s Technology Innovation Institute has made significant progress with Falcon-H1 Arabic, but a broader regional effort covering Egyptian Arabic, Gulf Arabic, Maghreb Arabic, and Levantine Arabic with the same fluency as English would be transformative for hundreds of millions of people. Such a model would require a collaborative effort across the Arab world, pooling data from diverse dialects and building evaluation frameworks that reflect the full spectrum of Arabic language use.

For Africa, the challenge is even more acute. With over 2,000 languages, the continent’s linguistic diversity dwarfs any other region’s. The Masakhane research community — now a network of more than 2,000 African researchers — established the Masakhane African Languages Hub in July 2025 and in January 2026 launched a major initiative to build AI datasets for 50 African languages, with the goal of empowering one billion Africans by 2029 through locally designed AI tools. The initiative, supported by Google.org, FCDO, IDRC, and the Gates Foundation, is building datasets for automatic speech recognition, real-world AI benchmarks, and culturally relevant multimodal data across 40 languages. But the gap between these foundational efforts and a Latam-GPT-scale deployment remains substantial.

The “AI for the Rest of the World” Movement

Latam-GPT is part of a growing global movement that challenges the assumption that AI must be built in a handful of wealthy countries and consumed everywhere else. This movement includes Cohere’s Tiny Aya models — released in February 2026 with support for over 70 languages and regional variants for African, South Asian, and Asia-Pacific languages — India’s Bhashini platform now running entirely on Indian cloud and GPU infrastructure, and the UAE’s Falcon model family, among many others.

What these projects share is a recognition that AI is not culturally neutral infrastructure like electricity or plumbing. AI systems encode language, values, and knowledge. When those systems are built exclusively by and for English-speaking Western societies, they inevitably marginalize the majority of the world’s population. The sovereign AI movement is, at its core, an assertion that every society has the right to AI that reflects its own language, culture, and priorities.

The economic argument reinforces the cultural one. As AI becomes embedded in education, healthcare, government services, and commerce, societies that depend entirely on foreign AI systems face a new form of digital dependency. They become consumers of a technology they do not control, subject to the pricing, policies, and priorities of foreign companies. Building local AI capacity is not just a matter of cultural pride — it is a matter of economic sovereignty.

Latam-GPT has proven that regional sovereign AI is technically feasible and organizationally achievable — and at a fraction of the cost that many assumed was required. The next question is whether other regions will follow Chile’s lead — and whether the global AI ecosystem will evolve into a multipolar landscape where diverse communities build AI that serves their own needs, or remain a unipolar system dominated by a handful of companies in a handful of countries.

Sources & Further Reading

• What Is Latam-GPT: Latin America’s Spanish and Portuguese AI Model — Euronews
• Latam-GPT and the Search for AI Sovereignty — Brookings Institution
• Launch of the First Open Large Language Model for Latin America & the Caribbean — Access Partnership
• Chile Launches Open Source Latam-GPT — Open Source For You
• Meet Latam-GPT, the New Open Source AI Model for Latin America — AI Business
• Chile Launches Latam-GPT in Push for Regional AI Sovereignty — bne IntelliNews
• Masakhane Hub Launches Funding Initiative for 50 African Languages — HapaKenya
• Falcon-H1 Arabic: World’s Leading Arabic AI Model — TII