1@misc{mineru,
2 title = {MinerU: A One-stop, Open-source, High-quality Data Extraction Tool},
3 author = {OpenDataLab},
4 year = {2024},
5 url = {https://github.com/opendatalab/MinerU}
6}
English summary
This repository bundles exactly the model files the mlx-mineru engine loads at inference time — a native C++/MLX reimplementation of MinerU for Apple Silicon, with zero Python at runtime. It is a model distribution, not a standalone runnable model.
Components: a 1.2B MinerU2.5 VLM (Qwen2-VL, safetensors, loaded directly by MLX) for the vlm/hybrid backends, plus ONNX models for the pipeline backend — PP-DocLayoutV2 (layout), PP-OCRv6 (DBNet detection + SVTR recognition), UniMERNet (formula), PP-LCNet (table classification), and SLANet+/UNet (table structure). The layout/OCR/formula ONNX were exported offline from the upstream PyTorch/Paddle models (scripts/export_*_onnx.py); the VLM weights are the original OpenDataLab safetensors.
Usage: place this folder as mumodel/ next to the mlx-mineru executable or in your working directory — it is auto-discovered (no --model flag needed).
License: mixed upstream licenses — the MinerU2.5 VLM is AGPL-3.0; the PaddleOCR/UniMERNet/PDF-Extract-Kit ONNX are Apache-2.0. Because the bundle includes AGPL-3.0 weights, treat the whole distribution as AGPL-3.0 and verify each upstream's terms before commercial/closed use. No upstream weights are re-licensed here.