---
title: README
emoji: 🌬️
colorFrom: blue
colorTo: indigo
sdk: static
pinned: false
---

# WindyWord

The models in this organisation are the copies used by the [Windy Word](https://windyword.ai) apps, built by [Windstorm Labs](https://windstormlabs.com) ([Windstorm Institute](https://windstorminstitute.org)).

**For developers, the canonical organisation is [WindyTranslate](https://huggingface.co/WindyTranslate)**, with scores, licences and attribution for every model at **[windytranslate.com](https://windytranslate.com)**.

## What is here

- 1,665 translation models (`translate-*`), derived from [OPUS-MT](https://huggingface.co/Helsinki-NLP), [HPLT](https://huggingface.co/HPLT), M2M-100 and MADLAD-400. Many repositories hold several deployment formats in subfolders (transformers and CTranslate2 INT8).
- 66 speech-recognition models (`listen-windy-*`, `windy-*-ggml`), derived from OpenAI Whisper, Distil-Whisper and community Whisper fine-tunes.
- [`WindyWord/WindyWord`](https://huggingface.co/WindyWord/WindyWord): an archived catalogue page from May 2026.

## Scores

Translation screening scores (chrF++ and BLEU on 48 FLORES-200 dev sentences) are published on [windytranslate.com](https://windytranslate.com) and on the WindyTranslate cards. Where a card here has a canonical copy in WindyTranslate, it links to it.

## Licensing

Every model keeps its upstream licence (Apache-2.0, CC-BY-4.0 or MIT) and carries the licence text and an attribution notice. Two groups are marked **licence under review** and are not for commercial use until the notice is removed: the WindyScripture variants (`herm0-scripture/`, trained on the eBible corpus, parts of which are non-commercial / no-derivatives) and five speech models whose upstream declares no licence.