---
tags:
  - ocr
  - multilingual
  - onnx
  - scriptmoe
---

# openocr-mlt ONNX Models

FP32 inference models for multilingual OCR with the `openocr-mlt` Python package. The bundle combines ScriptMoE autoregressive recognition with PP-OCRv6 medium text detection and provides whole-page OCR, text detection, and cropped text recognition.

## Model details

| Property | Value |
|---|---|
| Recognition | ScriptMoE ONNX Prefill/Decode with KV caching |
| Detection | PP-OCRv6 medium ONNX |
| Precision | FP32 |
| Effective batch size | 1 |
| Maximum recognition length | 100 decoding steps |

The archive includes recognition graphs, shared weights, the character dictionary, the model manifest, and the detector.

## Installation and use

Use Linux x86_64 with Python 3.10. Create a virtual environment and install the package.

~~~bash
python3.10 -m venv .venv-ocr
source .venv-ocr/bin/activate
python -m pip install openocr-mlt
~~~

The same installation supports CPU and NVIDIA CUDA on Linux x86_64. CUDA requires a compatible NVIDIA driver. The Python task API defaults to `device="auto"`. Set `device="cpu"` or `device="cuda:0"` to select a device explicitly.

~~~python
from openocr_mlt import OpenOCRMLT

with OpenOCRMLT(task="ocr") as ocr:
    result = ocr(image_path="page.jpg")
    print(result.text)
    print(ocr.info["device"])

with OpenOCRMLT(task="det", device="cpu") as detector:
    boxes = detector(image_path="page.jpg")
    print([box.points for box in boxes])

with OpenOCRMLT(task="rec", device="cpu") as recognizer:
    lines = recognizer(image_path="crop.jpg")
    print([(line.text, line.score) for line in lines])
~~~

Replace the image paths with existing files. Use `image=` for a PIL image or a uint8 NumPy array. NumPy arrays use BGR by default. Specify `color_order="RGB"` for RGB arrays.

The first task call downloads and verifies the registered model bundle. ModelScope is the primary source, with Hugging Face and hf-mirror as fallbacks. The default model cache is `$XDG_CACHE_HOME/openocr-mlt` or `~/.cache/openocr-mlt`.

First GPU use may require about 1.32 GB of additional CUDA 12 and cuDNN 9 libraries. The library preparation process displays download and verification progress. Compatible local libraries and verified caches can be reused. PyTorch is optional.

## Model archive

| Property | Value |
|---|---|
| File | `scriptmoe-medium-fp32-20261101-v1.tar.gz` |
| Size | 213,794,128 bytes |
| SHA256 | `ee43e7358066ddf1286f61983024dc407808e350b1dd57eb986523eaf9417ad3` |

For offline use, copy the archive to the target machine and import it with the installed package.

~~~bash
python -m openocr_mlt fetch-models \
  --source ./scriptmoe-medium-fp32-20261101-v1.tar.gz \
  --cache-dir ./models --offline
~~~

The command verifies the archive and model files before installing them in the cache. In the Python API, set `cache_dir="./models", offline=True` to use this model cache offline. CUDA use also requires compatible local runtime libraries or a previously prepared CUDA cache. Offline package installation requires local copies of the package dependencies.

## Sources and licensing

Recognition derives from [ScriptMoE](https://github.com/YesianRohn/ScriptMoE), upstream revision `5dde99867aaee6b4af3e75a3020b225eab17ac53`. Detection uses `PP-OCRv6_medium_det_onnx.tar` from Paddle's official model storage. Component attribution and licensing details are provided in [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).

Package installation and API documentation are available on [PyPI](https://pypi.org/project/openocr-mlt/).