Tucano

Tucano is a family of open decoder-only language models built specifically for Portuguese, developed by researchers at the University of Bonn. The project started from the data problem: Portuguese is a low-resource language in AI terms, so the team assembled GigaVerbo, a deduplicated 200-billion-token Portuguese corpus, and trained a range of model sizes on it, including a 2.4B variant. Models, datasets, and tooling are all openly released for reproducible research.

In independent Brazilian benchmarks of Portuguese generation, Tucano variants consistently outperform comparable multilingual models of similar size on summarization and simplification tasks. The work is part of a small but real wave of Brazilian and European Portuguese modeling alongside projects like Bode and Sabia, and it gives Portuguese-first applications a fully open foundation instead of a translated afterthought.