Lumi launches with 4 billion parameters and runs without a data center GPU

Lumi is the language model released by the Centro de Excelência em Inteligência Artificial (AI Center of Excellence), a research center at Universidade Federal de Goiás (a federal university in Brazil's Goiás state). The announcement came on Wednesday, October 28.

The project targets the Portuguese language and the Brazilian context.

The size explains much of the interest. However, it has 4 billion parameters, a number that balances performance and operational efficiency.

As a result, the model runs on more accessible computing setups. Moreover, large infrastructure investments are no longer a requirement.

Where to find it and start testing

Distribution follows the expected path for an open project.

The model will be made available on the Hugging Face platform. However, free use applies to the development of new solutions.

In addition, materials related to the project appear on GitHub.

Therefore, the starting point is well documented for those who want to evaluate it before integrating it.

Lumi targets dependence on foreign technology

This point stands out as the project's central motivation.

Arlindo Galvão, Director of Open Innovation at the center, summarized the diagnosis in a statement. According to him, many Brazilian companies depend on solutions created abroad that do not always address the particularities of the local market.

The stated goal is autonomy. The objective involves customizing the technology to specific needs and reducing entry barriers.

The four practical advantages of an open model

Being open source brings very concrete consequences.

First, it allows incorporating data specific to the user's activities. This happens without sharing sensitive information with third parties.

Second, companies gain freedom to build applications with local infrastructure. This way, dependence on services billed in foreign currency decreases.

Third, there is the possibility of customization through fine-tuning. The technique submits the already trained model to a smaller, more specific dataset.

Fourth, the same base gives rise to varied applications. Different segments, therefore, make use of the same starting point.

The context that favors the launch

The project arrives at a specific moment for the country.

The Plano Brasileiro de IA (Brazil's national artificial intelligence plan) calls for R$23 billion in investments over the coming years.

In addition, recent official statements reinforced this direction. Attracting data centers and launching a Brazilian AI were among the proposals for the sector.

What to assess before moving to production

First, compare the total cost against commercial APIs. Self-hosting trades a dollar-denominated monthly fee for infrastructure and maintenance.

Second, size your expectations to the model's scale. A 4-billion-parameter model handles specific tasks well after fine-tuning.

Third, test with your own data before deciding. A generic benchmark says little about your use case.

Finally, consider the privacy gain. Processing sensitive data on your own infrastructure meets requirements that no external API can fulfill.

Follow our profile on Instagram!