UFG’s CEIA Launches Lumi, a Free 4-Billion-Parameter Portuguese Model That Companies Can Train on Their Own Data
The Centre of Excellence in Artificial Intelligence (Ceia) at the Federal University of Goiás (UFG) has released CEIA Lumi, an open language model built for Portuguese and the Brazilian context, Jornal Opção reported. The model was launched on Wednesday, October 28, at the AI Brasil Experience 2026 event in São Paulo and is free to download, with the aim of letting companies use it as a base for their own tools.
Lumi has 4 billion parameters, a size chosen so it can run on more affordable computing setups. Unlike chat services such as ChatGPT or Gemini, it is meant to be downloaded, fed with an organisation’s own data and trained further for specific uses, including on a company’s own infrastructure. Arlindo Galvão, Ceia’s director of open innovation and one of the developers, said dependence on foreign platforms and operating costs often tied to the dollar keep many Brazilian businesses away from AI.
According to Ceia, the model was trained on more than 100 billion tokens of Portuguese text to better handle expressions and features of Brazilian usage. Galvão said the centre has already tested the approach in sectors such as energy and education, including a project with the utility Cemig that produced a power-sector model called Energy GTT. He said Lumi performed competitively against general-purpose models in tests, while noting that results depend on the task.
What it means in practice
Lumi joins a small but growing group of Brazilian-built language models. The national computing laboratory recently released Manacá-1B, an open Portuguese model, and large private players are building their own, as when Itaú agreed to buy NeoSpace to develop in-house foundation models. A free model from a public university gives smaller companies a starting point without a large licence bill.
The practical limits are real. Firms still need people who can configure and retrain the model, and a 4-billion-parameter system will not match the largest commercial models on every task. Running it locally can help keep sensitive data inside a company, but only if the surrounding systems are designed for that. Galvão said the centre had no count yet of interested firms, although it is already in contact with businesses in Goiás.
Ceia director Telma Soares said the launch is meant to bring scientific research closer to market needs. Lumi is available on Hugging Face, with support material due on GitHub. What to watch next is the specialised versions Ceia says it is preparing for health, law and education, and whether Brazilian companies publish real deployments built on the model.