Research
2026 ranking of Brazil's AI labs, by research and global reach
Who competes on international leaderboards, who published new research this year and who is most cited and downloaded worldwide. Five labs, with the rule and the sources in the open.
Brazil does AI research in companies, universities and public centers. To compare them we used three public signals, measured on November 6, 2026: a published result on an international frontier benchmark in 2026, a new study or model published in 2026 with method and numbers, and accumulated global reach, through OpenAlex citations and Hugging Face downloads over the last month.
The order follows the three signals in that sequence: first who competes today on the same benchmarks as labs in the US and China, then who published new research this year, then accumulated reach.
The ranking
| Lab | International benchmark in 2026 | New study or model in 2026 | Citations | Downloads, last 30 days |
|---|---|---|---|---|
| 1. Rogue | 3 benchmarks: Terminal-Bench 4.0, TB-Science 0.1, DeepSWE v1.1 | 3 studies and open data | 0 | 0 |
| 2. Maritaca AI | Evaluation in Portuguese, in its technical report | Sabiá-4 report (March) and S5 (November) | 145 | 583 |
| 3. Unicamp and NeuralMind | None found | None found | 903 | 1,088,467 |
| 4. Recogna (Unesp) | None found | None found | 31 | 3,543 |
| 5. CEIA-UFG | None found | None found (GAIA is from June 2025) | 0 | 1,598 |
Lab by lab
1. Rogue. The Brazilian lab with published results on international agent benchmarks in 2026: 71.21% on Terminal-Bench 4.0, ahead of Claude Opus 5.5 (66.4%), 71.43% on Terminal-Bench-Science 0.1 and 75.22% on DeepSWE v1.1, ahead of GPT-6 Astra. It published three studies this year, with cost per task, and opened the DeepThink aggregate data.
2. Maritaca AI. Trains the Sabiá family for Portuguese and the Brazilian context, published the Sabiá-4 technical report in March 2026 and launched S5 on November 3. The family's reports add up to 145 citations. The company was founded by Unicamp researchers.
3. Unicamp and NeuralMind. BERTimbau, a Portuguese language model released in 2020, has 702 citations and about 1.08 million downloads in the last month. mMARCO and PTT5, from the same school, add 201 citations.
4. Recogna (Unesp). Published Bode, a model tuned for Portuguese tasks, with 31 citations, and keeps 45 open models on Hugging Face, with 3,543 downloads in the last month.
5. CEIA-UFG. UFG's AI center released GAIA in June 2025, a Portuguese version of Gemma 3 built with Google DeepMind and partners, with 1,598 downloads in the last month.
Who is left out
Broad academic centers, such as USP's C4AI with IBM and FAPESP and Unicamp's Recod.ai with about 200 researchers, publish across many areas of AI and do not fit a ranking of models and agents. Companies whose research is not published in a measurable form, such as WideLabs, maker of Amazônia IA, are also left out.
How to redo the math
International benchmark: Terminal-Bench 4.0, Terminal-Bench-Science 0.1 and DeepSWE v1.1. Research in 2026: a report, paper or study with method and numbers published between January and November 6, 2026. Citations: OpenAlex sum of each lab's main papers. Downloads: the lab's Hugging Face organization over the last 30 days, read on November 6, 2026.
