Language model
Salamandra-7B, from the Barcelona Supercomputing Center, fine-tuned for normative Galician.

The normative Galician proofreader that runs on your own computer, no cloud.
Your texts never leave your device: Peneira runs locally, sending nothing to any server. And, above all, it is transparent — it explains every correction and says where it comes from. It suggests, it doesn't impose.
Beta · one-click download · works offline
Peneira detects and corrects Castilian interference and normative errors in Galician texts, grounded in the DRAG and the RAG/ILG standard. It is meant for anyone who writes in Galician and wants normative confidence with a tool that explains, rather than correcting blindly.
It combines the flexibility of a language model with the safety of deterministic rules.
Salamandra-7B, from the Barcelona Supercomputing Center, fine-tuned for normative Galician.
Rules that guarantee clear Castilianisms, official place names, plurals, the te/che regime, pronoun placement and more never slip through.
It tells Galician from Castilian and warns you, instead of half-correcting.
Every correction is tagged with its origin (model or rule), so you know why it is suggested.
The beta ships with its numbers and its limits in plain sight.
The downloadable version runs 100% on your computer: texts are never sent to any server.
The demo on this page runs in your browser (it sends the text nowhere) and is only illustrative; the real, full version is the downloadable one.
An interactive demonstration of what Peneira does. Type —or use the example— and see the corrections.
Illustrative simulation (real sample corrections). The full proofreader is the downloadable version.
One-click download. Unzip and double-click Peneira.bat.
Download Peneira.zipRequirements: Windows · ~8 GB RAM recommended · space for the unzipped files.
Peneira openly publishes its evaluation resources —the test set, the normative inventory and the architecture documentation— under a CC BY 4.0 licence. It is measured on a closed test set, with detection around 96%.
Developed by Argalla. Base model Salamandra (BSC, Apache-2.0) · Real Academia Galega / ILG standard · Nomenclátor 2026 place names (CC BY-SA).
Contact and error reports: info@argalla.com