Home / Models / translate-windy-max

Multilingual translation model translate-windy-max

One CTranslate2 INT8 model for many language directions, fine-tuned by Windstorm Labs from google/madlad400-3b-mt.

Task: Translation (multilingual)
Lineage: MADLAD-400
Format: CTranslate2 (INT8)
Hugging Face: WindyWord/translate-windy-max
Last updated on HF: 2026-07-27

Quality (as published in the model card)

FLORES-200 devtest, 1012 sentences per pair, beam 4. These numbers come from the Hugging Face model card; this site has not re-measured them, and they use a different setup from our pair screening (devtest, 1,012 sentences, spBLEU and chrF), so do not compare them directly with pair scores.

Pair	spBLEU	chrF
English → Spanish	32.77	56.33
English → French	55.78	71.93
English → German	47.48	66.73
English → Italian	37.27	60.01
English → Portuguese	54.49	71.54
English → Russian	40.48	59.22
English → Chinese	33.79	34.93
English → Japanese	25.03	34.89
English → Korean	26.21	35.86
English → Arabic	39.22	57.40
English → Hindi	34.84	54.88
English → Swahili	30.81	56.18
Spanish → English	35.37	60.72
French → English	49.78	69.64
Chinese → English	32.57	58.13
Japanese → English	30.98	56.95
Mean	37.93	56.58

Covers 76 of the 76 languages in the Windy translation set. Full coverage. Tagalog uses the `<2fil>` tag.

How to run

Shell

pip install ctranslate2 transformers sentencepiece huggingface_hub

Python

import ctranslate2
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer

path = snapshot_download("WindyWord/translate-windy-max")
tok = AutoTokenizer.from_pretrained(path)   # the tokenizer ships in the repo
translator = ctranslate2.Translator(path, device="cpu", compute_type="int8")

# The target language is a tag at the start of the input: <2es> = Spanish.
source = tok.convert_ids_to_tokens(tok.encode("<2es> Where is the train station?"))
result = translator.translate_batch([source], beam_size=4)
print(tok.decode(tok.convert_tokens_to_ids(result[0].hypotheses[0]), skip_special_tokens=True))

Tested: Pattern run on 2026-10-01 against WindyWord/translate-windy-max with ctranslate2 4.8.2 and transformers 5.18.0 on CPU. Not run on every model individually.

Licence and attribution

Licence: Apache-2.0
Upstream: google/madlad400-3b-mt
Upstream licence: Apache-2.0
Attribution: Based on google/madlad400-3b-mt (MADLAD-400) by Google. Licensed Apache-2.0.

Copy this into your NOTICE file (or equivalent) when you ship the model:

NOTICE

This product uses the model WindyWord/translate-windy-max
(https://huggingface.co/WindyWord/translate-windy-max), published by Windstorm Labs (https://windstormlabs.com).
Based on google/madlad400-3b-mt (MADLAD-400) by Google. Licensed Apache-2.0.
Windstorm Labs modified the upstream weights; the model card describes what was changed.
Licensed under the Apache License, Version 2.0: https://www.apache.org/licenses/LICENSE-2.0

This is a convenience, not legal advice. See Licensing.