Introducing Latam-GPT: The New Open Source AI Model Revolutionizing Latin America

Key Takeaways

Chile has launched Latam-GPT, an open-source AI model tailored for Latin American languages and cultures.
Developed by over 60 institutions from 15 countries, the model aims to enhance regional technological sovereignty.
Despite limited funding of $550,000, Latam-GPT is now available on Hugging Face and GitHub for further development.

Introduction of Latam-GPT

Chile has unveiled a new open-source AI model named Latam-GPT, specifically trained on the languages and cultural contexts of Latin America. This innovative project has been in development for over two years, led by the National Center of Artificial Intelligence (CENIA) in collaboration with more than 60 institutions from 15 different Latin American and Caribbean nations. Key partners include the Chilean Ministry of Science, Technology, Knowledge and Innovation, AWS, and the Development Bank of Latin America and the Caribbean.

Focus on Regional Context

Latam-GPT is designed to address concerns over the global dominance of AI technologies by major U.S. tech companies. According to CENIA, the model captures cultural and linguistic nuances absent in AI systems trained mainly on English and northern global frameworks. “Latam-GPT allows Latin America to join the AI revolution as a key player,” stated CENIA director Alvaro Soto at its launch event in Santiago on February 10. Science Minister Aldo Valle emphasized the importance of regional integration for achieving technological sovereignty.

Data Representation Challenges

There is an evident need for technologies like Latam-GPT, especially considering that research indicates Spanish—spoken by the majority of Latinos—comprises only about 4% of the data used to train existing language models, while Portuguese accounts for as little as 2%. The new model utilizes both Spanish and Portuguese as its primary languages and aims to incorporate indigenous languages in future developments.

Technical Specifications

Latam-GPT is based on the Meta Llama 3.1 open model, boasting 70 billion parameters. It has been trained on over 300 billion plain-text tokens—approximately 230 billion words—culled from officially sanctioned texts. This extensive dataset has been curated to provide what CENIA claims is a “high quality dataset.”

Market Positioning and Future Prospects

Despite its promising features, Latam-GPT faces challenges in penetrating the highly competitive AI market, which is dominated by a few major companies from the U.S. and China. The project was developed on a modest budget of $550,000, which may constrain its immediate market impact. However, the model’s availability on platforms like Hugging Face and GitHub suggests that it may serve as a foundational tool for developers seeking to create region-specific applications in the future.

The content above is a summary. For more details, see the source article.