Download, self-host, benchmarked. From Windstorm Labs.
Every model is public on Hugging Face. No key, no account, no network call after the download.
Windy Translatepip install "transformers<5" sentencepiece torchfrom transformers import pipeline
translate = pipeline("translation", model="WindyTranslate/translate-tc-big-en-fr")
print(translate("Where is the train station?")[0]["translation_text"])
# Où se trouve la gare ?The translation pipeline was removed in transformers 5. This version works on both 4.x and 5.x and is the pattern every model page uses:
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
repo = "WindyTranslate/translate-tc-big-en-fr"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSeq2SeqLM.from_pretrained(repo)
batch = tok(["Where is the train station?"], return_tensors="pt", padding=True)
print(tok.batch_decode(model.generate(**batch), skip_special_tokens=True)[0])
# Où se trouve la gare ?Both run on CPU. Output shown is what we got on 2026-10-01. Each model page has the right snippet for its format (transformers, CTranslate2, faster-whisper or whisper.cpp) and says what was tested.
Or from Python: pip install windytranslate, then windytranslate find en sw picks the best measured model for a pair and prints its licence and attribution (PyPI, source).
More on GitHub: windytranslate-examples: find the best model for a pair from models.json, translate locally sentence by sentence, and self-host one model behind HTTP with Docker (CPU, offline, ships its NOTICE).
788 pair models scored on FLORES-200 dev (48 sentences each, chrF++). Median chrF++ 36.79. 787 more are not yet scored, mostly because FLORES-200 has no test set for their languages; 53 of those passed a GlotLID language-identity check.
Screening scores, not publication results. How we measured, and the caveats.
| Model | chrF++ (FLORES-200 dev) | Licence |
|---|---|---|
| English to French translate-tc-big-en-fr | 69.6 Excellent | CC-BY-4.0 |
| Bulgarian to English translate-tc-big-bg-en | 68.8 Excellent | CC-BY-4.0 |
| Swedish to English translate-sv-en | 67.8 Excellent | Apache-2.0 |
| English to Portuguese translate-tc-big-en-pt | 66.6 Excellent | CC-BY-4.0 |
| English to Bulgarian translate-tc-big-en-bg | 66.6 Excellent | CC-BY-4.0 |
| Danish to English translate-da-en | 66.3 Excellent | Apache-2.0 |
| English to French translate-en-fr | 66.2 Excellent | Apache-2.0 |
| French to English translate-tc-big-fr-en | 66.2 Excellent | CC-BY-4.0 |
Recorded output from our benchmark log, not written by hand. Same FLORES-200 sentence across pairs, so you can compare. Includes a long-tail pair and a pair we do not recommend.
Real output, FLORES-200 dev, recorded 2026-08-04. The log keeps the first 110 characters of sentence 1 of each pair's run; the score is over all 48 sentences.
Windstorm Labs is the engineering arm of the Windstorm Institute. We fork from strong open foundations (OPUS-MT, HPLT, M2M-100, MADLAD-400, Whisper), refine them, and publish every score, including the weak ones. Windy Translate is where developers get those models.
models under care in the lab's model clinic, each with a chart of where it came from, what was changed and what the checks found.
builds maintained across the lab: base weights, fine-tuned variants and quantized formats.
Clinic figures are from the lab's internal roster; catalogue figures are computed from models.json.
Windy Traveler is the consumer app: the polished storefront for travellers. Windy Translate is the warehouse behind it, for developers.
To be exact about today: since 2026-10-01, Traveler's typed translations between English and Spanish, French, German or Portuguese (both ways) run on pairs from this catalogue, on Windy's own servers. We measured the swap on the full FLORES-200 devtest (1,012 sentences per direction): each of the eight models scores within 1 chrF++ of the engine it replaced, five of them higher, and all are Apache-2.0 or CC-BY-4.0. Other languages, Chinese, Japanese and voice are handled by other services; Traveler's FAQ says which.
The hosted API is not live yet. It will be free while in beta; pricing after the beta has not been set, and we will announce it before anything changes.