iPaklink on-device model bundle
Offline model files for the iPaklink Android app. Translation, OCR and speech recognition all run locally on the device; no server-side inference and no network connection is required after installation.
Contents
| Group | Model | Files | Size |
|---|---|---|---|
| Translation | m2m100_fourlang int8 (12 directions across zh / en / ru / uz) |
1 | 734.1 MiB |
| OCR | PP-OCRv6 tiny (det / rec / Cyrillic rec) | 5 | 14.6 MiB |
| ASR | sherpa-onnx zh / en / uz + Vosk ru | 25 | 273.9 MiB |
| Total | 31 | 1,072,146,944 bytes (1.0 GiB) |
The directory layout is identical to the app's OSS object keys, e.g.
models/asr/v1/uz/tokens.txt.
Downloading a pinned version
Always address files by the 40-character commit SHA, never by main:
https://huggingface.co/wingyouth/on-device-models/resolve/<COMMIT_SHA>/models/asr/v1/uz/tokens.txt
manifest.json lists every file with its exact sizeBytes and sha256;
SHA256SUMS is provided for bulk verification. Verify the hash after download:
a mirror may occasionally ignore an HTTP Range request and return the whole
file with HTTP 200, producing a file of the right size but wrong content.
Runtime requirements
- onnxruntime-android 1.17.1 (sherpa-onnx native library depends on
OrtGetApiBase@@VERS_1.17.1; other versions fail at load time). - ONNX IR version <= 9.
- Model files are downloaded at first use and are not bundled in the APK.
License
iPaklink's own contributions are MIT (see LICENSE). Upstream weights keep
their own licenses - see NOTICE. The upstream license inventory is still
pending verification; do not rely on this repository for commercial
redistribution until it is complete.
Status
Published by cmd/modelpush/publish_hf.py. This repository currently serves
the Uzbekistan / non-mainland-China distribution side only.