# Open Portuguese‑Nheengatu translation models released

> Researchers publish the largest Portuguese‑Nheengatu corpus and open‑weight MT models, reaching up to 32.8 chrF++ on legal text.

Oossa · 2026-10-14 · https://oossa.com/en/open-portuguese-nheengatu-translation-models-released

Fabrício Carraro and co‑authors released open‑weight machine‑translation models for Portuguese‑Nheengatu, the Amazonian Língua Geral spoken by 6‑10 k people. The models are trained on a new 14.5 k sentence parallel corpus that includes the 2023 Portuguese translation of Brazil’s Constitution. They also added a speech‑derived test set, which lifted the speech translation score by about 25 chrF++ points. The best Portuguese‑to‑Nheengatu system scores 31.7‑32.8 chrF++ on unseen constitutional articles.

## The facts

- 14.5 k Portuguese‑Nheengatu sentence pairs collected
- Best Portuguese‑to‑Nheengatu chrF++ score: 31.7‑32.8 on disjoint constitutional text

## Why it matters

The open models give speakers of Nheengatu a usable tool for legal and everyday communication in their language.

## Sources & references

1. [The Constitution Speaks Nheengatu: An Open MT System and Reproducible Corpus Pipeline for the Amazonian L\'ingua Geral](https://arxiv.org/abs/2610.14760) – arXiv, 2026-10-14

Last updated: 2026-10-14