Introducing GAIA, a Brazilian Portuguese Gemma 3 model developed with ABRIA, CEIA, Nama, and Amadeus AI The Brazilian Association of AI (ABRIA), Center of Excellence in Artificial Intelligence (CEIA) at Universidade Federal de Goiás under the coordination of Professors Celso Camilo and Sávio Teles, startups Nama and Amadeus AI, and Google DeepMind have partnered to develop GAIA, a new open model based on Gemma 3 to advance Brazilian Portuguese language understanding.The goal of releasing the open model is to enable developers and organizations to build on credible foundational technology, lowering the barriers of adoption for AI-powered solutions.The challengeWhile general-purpose models offer broad capabilities, they often lack the linguistic and cultural understanding needed for localized content, education, customer service, and other common applications due to underrepresentation in their initial training datasets.This can lead to less accurate translations, culturally insensitive outputs, and failure to capture the full context of conversations in Brazilian Portuguese. Consequently, applications built on general-purpose models may underperform.The solutionBased on the Gemma 3 4B multimodal model, GAIA was developed through continuous pre-training with a large set of Portuguese-language data. The proportion of samples from each source were dynamically adjusted during training to ensure a balanced mix of different types of data.Training was carried out on an NVIDIA DGX (Deep GPU Xceleration) infrastructure with H100 GPUs, in total processing approximately 13 billion tokens. To restore the model’s ability to follow instructions without resorting to traditional fine-tuning, the instruction residuals technique as described in “Balancing Continuous Pre-Training and Instruction Fine-Tuning: Optimizing Instruction-Following in LLMs” was applied for mapping operations.In internal evaluations, covering questions from the ENEM Brazilian national standardized exam and the ASSIN2 benchmark for semantic similarity and textual inference in Portuguese, the model demonstrated superior performance even without any dedicated fine-tuning for instructions.The impactBy establishing an open foundation model, organizations and individual developers can further adapt GAIA and build solutions tailored for conversational AI, language translation, text generation, and other tasks in Brazilian Portuguese.The Tribunal de Contas dos Municipios de Goiás (TCM-GO) is exploring GAIA in ChatDoc, a solution for simple and complex document scenarios, and has observed circumstantial characteristics indicating a more robust question interpretation behavior compared to the original gemma3-4b-it model.Marcelo de Oliviera, Technology Manager at TCM-GO, notes, “The GAIA model has a characteristic where it explains the question’s context in greater detail without straying from the main context, not adding unrelated details, which is often observed in the original model.”In addition, the 4B size enables execution on proprietary hardware, reducing API costs.Other organizations that plan to test GAIA in applications include:Tribunal de Contas do Estado de Goiás (TCE-GO) to summarize content and detect similarity among organizational processesUnimed Fesp to validate discopathy diagnoses in medical records for insuranceBHub for customer service applicationsBeNext for the development of conversational chatbotsRecife Center for Advanced Studies and Systems (CESAR) for a tool to help public school teachers score essays from students.What’s nextGAIA is now available on Hugging Face. From research centers and NGOs to startup companies, anyone can use this Portuguese-driven model.