RapidOCR model mirror
A mirror of the RapidAI/RapidOCR
model repository, re-hosted on Hugging Face so that downstream services can
fetch the weights from Hugging Face instead of modelscope.cn.
The files here are byte-for-byte copies of the corresponding files in the
upstream ModelScope repository at tag v3.9.2 (302 files, ~3.6 GB: ONNX,
MNN, Paddle, and Torch model formats, character dictionaries, fonts, and test
files). No files were modified; relative paths are preserved.
URL mapping
A file at the upstream revision URL
https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/v3.9.2/onnx/PP-OCRv6/rec/PP-OCRv6_rec_small.onnx
is mirrored at
https://huggingface.co/DjB314/RapidOCR/resolve/main/v3.9.2/onnx/PP-OCRv6/rec/PP-OCRv6_rec_small.onnx
In other words, the upstream revision prefix resolve/v3.9.2/… maps to
resolve/main/v3.9.2/… here, and the directory layout below the version
matches upstream exactly. Any tool that swaps the host + repo can therefore
use this mirror without changing paths.
Usage
Direct download:
curl -L -o PP-OCRv6_rec_small.onnx \
https://huggingface.co/DjB314/RapidOCR/resolve/main/v3.9.2/onnx/PP-OCRv6/rec/PP-OCRv6_rec_small.onnx
Hugging Face Hub:
hf download DjB314/RapidOCR v3.9.2/onnx/PP-OCRv6/rec/PP-OCRv6_rec_small.onnx --local-dir .
# or the whole tree:
hf download DjB314/RapidOCR --repo-type model --local-dir rapidocr-models
Downstream tools that previously fetched from the upstream ModelScope URL can switch to this mirror by swapping the host+repo in the URL — see the URL mapping section. The relative paths below the version are identical to upstream, so no path changes are needed.
The files carry no modification timestamps from upstream, so integrity checking is by checksum: see verification.
Versioning & refresh policy
| Revision mirrored | Files | Size | Date |
|---|---|---|---|
v3.9.2 |
302 | ~3.6 GB | 2026-08 |
This mirror tracks upstream tags only (never arbitrary main changes).
When RapidAI tags a new release, the mirror is re-synced and this table is
updated. If you depend on a newer tag before that happens, open an issue.
Verification
Each staged file is copied verbatim from upstream and checked against the SHA-256 recorded in the upstream ModelScope file listing; mismatches are re-downloaded or reported, never shipped. To verify a downloaded file locally:
sha256sum PP-OCRv6_rec_small.onnx
# compare against the upstream listing:
# https://www.modelscope.cn/api/v1/models/RapidAI/RapidOCR/repo/files?Revision=v3.9.2&Recursive=true
License & attribution
Licensed under the Apache License, Version 2.0 — the same license declared by the upstream sources. See LICENSE.
- RapidOCR — github.com/RapidAI/RapidOCR and the model repository modelscope.cn/models/RapidAI/RapidOCR (Apache-2.0). Original source of the files mirrored here.
- PaddleOCR (PP-OCR) — github.com/PaddlePaddle/PaddleOCR (Apache-2.0). Upstream origin of the PP-OCR model weights.
This mirror redistributes the above works unmodified, retaining their Apache-2.0 license and attribution as required by Section 4 of the license.
This is an independent mirror. It is not affiliated with, endorsed by, or sponsored by RapidAI, ModelScope, or PaddlePaddle. Product names and marks are used for identification only.