Add model card and attribution

README.md

@@ -0,0 +1,37 @@
+---
+license: cc-by-4.0
+language:
+- zh
+- en
+pipeline_tag: translation
+tags:
+- ctranslate2
+- int8
+- android
+- translation
+---
+
+# QuickMT INT8 Android
+
+This repository contains on-disk INT8 CTranslate2 derivatives of the QuickMT Chinese-English translation models for Android CPU inference.
+
+Upstream models:
+- zh-en: https://huggingface.co/quickmt/quickmt-zh-en at revision `c27cc8024e01a047733a1e34796e2ab19d74b237`
+- en-zh: https://huggingface.co/quickmt/quickmt-en-zh at revision `f48b69e53866b717607b59f94350e9aae3ed021c`
+
+The upstream models are licensed under CC BY 4.0. These derivative weights preserve that license and attribution.
+
+## Transformation
+
+The original CTranslate2 floating-point weights were rewritten to on-disk INT8 using CTranslate2-compatible per-row symmetric weight quantization. Tokenizers, vocabularies, and model configuration are retained from the pinned upstream revisions.
+
+## Files
+
+- `zh-en/` - Chinese to English
+- `en-zh/` - English to Chinese
+
+Validated `model.bin` SHA-256:
+- zh-en: `0d2febd770738a14bfad064ebccc38a690615ebec29adf877132a4e61371dbe0`
+- en-zh: `83bd025a6d9ccf3869d924051263b4a7bce3d41ca1d26ccd788a5d337df6c58b`
+
+The quantized candidates were device-tested against the prior runtime models on an Android arm64 device. The project test set produced identical translations while reducing model storage and load-time overhead. This is not an official QuickMT release.