UFG launches Lumi, 4-billion-parameter open-source AI for Portuguese
The CEIA model was built for Brazilian reality, runs on more affordable infrastructure and will be free on Hugging Face.
In 30 seconds
- UFG's CEIA has launched Lumi, a language model built for the Portuguese language and Brazil's reality.
- With 4 billion parameters, it balances performance and efficiency and runs on more accessible infrastructure.
- As open source, it lets users work with their own data without sharing sensitive information and customize it with fine-tuning.
- The model will be free on Hugging Face, with related materials on GitHub.
On Wednesday (28), Lumi was unveiled by the Center of Excellence in Artificial Intelligence (CEIA), affiliated with the Federal University of Goiás (UFG). It is a language model focused on Portuguese and tuned to the Brazilian context. The initiative was developed to meet the domestic market's growing demand for AI tools tailored to the country.
What Lumi does in practice
Lumi has 4 billion parameters, the internal settings a model uses to understand and generate text. At this size, it aims for a middle ground between performance and operational efficiency, without requiring massive infrastructure to run. According to the developers, this allows it to be used on more accessible computing setups, without major investment.
The model is open source, meaning it can be freely used, studied and adapted. In practice, this means companies and developers can build their own tools with greater autonomy and flexibility. The goal is to expand Brazil's role in the global AI market and reduce dependence on foreign technologies.
Private data and customization
One of the key points is the ability to add data specific to the user's business without sharing sensitive information. This matters for companies handling internal information that they do not want to send to outside services. The announcement also highlights greater freedom to build applications with local infrastructure, which helps reduce reliance on services billed in foreign currency.
Another feature is customization through fine-tuning, a technique that takes an already trained model and trains it again with a smaller, more specific dataset. It is like giving the AI a specialization course so it learns the vocabulary of a field such as customer service, education or retail. That way, the same core technology can power varied applications for different sectors.
Today, many Brazilian companies depend on solutions developed abroad, which do not always meet the particular needs of our market. We want to offer an alternative that lets companies innovate with more autonomy, tailoring the technology to their needs and lowering barriers to entry
Arlindo Galvão, Director of Open Innovation at CEIA
Where to use it for free
The Brazilian model will be available for free on the Hugging Face platform, which works as a showcase where anyone can download and test AI models. Materials related to the project will also be available on GitHub, the repository where developers publish code and documentation. The goal is to make it easier to create new solutions focused on the Brazilian market.
Sources
The facts in this story come from these outlets:
- Estadão “Lumi: conheça o modelo de IA nacional desenvolvido pela UFG”
Text produced with the help of artificial intelligence from the sources cited and reviewed before publishing. The video presentation is AI-generated.
Add CodeMago News as a preferred source on Google


