PaddleOCR-VL
PaddleOCR-VL is PaddlePaddle's two-stage document parser. PP-DocLayoutV2 finds 25 kinds of layout region and their reading order, and PaddleOCR-VL-0.9B reads each region as text, a table, a formula or a chart. LibreYOLO ports the PaddleX pipeline around both networks for the ocr task.
- Tasks
- ocr
- Install
pip install "libreyolo[vlm]"- Support tier
- Sibling tier, since v1.8.0. A separate product surface with its own factory and contract.
- Licenses
- Code Apache-2.0, weights Apache-2.0. Commercial use
Install
PaddleOCR-VL belongs to LibreYOLO's VLM tier and needs the vlm extra. Both
networks load through native transformers classes, which needs transformers
5.3 or newer. PaddlePaddle itself is not needed.
pip install "libreyolo[vlm]"Predict
Weights for both networks download on first use from one pinned mirror,
LibreYOLO/LibrePaddleOCRVL0.9b. No remote code runs.
from libreyolo import LibreVLM model = LibreVLM("paddleocr-vl") # weights download on first useresult = model.predict("page.png") # a scanned or rendered page for label, text in zip(result.ocr.labels, result.ocr.texts): print(label, text[:60]) # one region per block, reading orderprint(result.ocr.markdown) # the whole page as markdownfrom libreyolo import LibreVLM # Skip layout detection and read a pre-cropped formula.model = LibreVLM("paddleocr-vl", use_layout_detection=False, prompt_label="formula")result = model.predict("formula.png")print(result.ocr.texts[0]) # LaTeXfrom libreyolo import LibreVLM # Charts are skipped by default, as upstream; opt in to read them as tables.model = LibreVLM("paddleocr-vl", use_chart_recognition=True)result = model.predict("report.png")The family loads through LibreVLM(), or the LibrePaddleOCRVL class.
Results.ocr holds one region per recognized block (a paragraph, a title, a
table), not one per text line, in reading order: texts, polygons
(axis-aligned boxes as 4-point quads), labels (the layout class) and the
page in markdown. Tables come back as HTML and formulas as LaTeX. Pictures,
and charts and seals by default, are left out of Results.ocr and the
markdown.
Results.ocr.conf is the geometric mean of the generated tokens'
probabilities and det_conf the layout score. conf= does not gate regions;
the layout keeps the upstream per-class thresholds, which layout_threshold=
overrides. Recognition is greedy with up to 4096 new tokens, and each region is
resized within min_pixels / max_pixels (default 112896 / 1003520).
use_layout_detection=False with prompt_label= (ocr, table, formula
or chart) reads a pre-cropped element.
Variants
One size, 0.9b: a NaViT-style vision encoder and the ERNIE-4.5-0.3B decoder,
behind PP-DocLayoutV2. CUDA runs the recognizer in bf16 (fp16 without bf16
support) and the layout detector in fp32; CPU and MPS run both in fp32. On MPS,
set PYTORCH_ENABLE_MPS_FALLBACK=1 if an operator is reported missing.
Not ported: adjacent text-block merging, figures embedded in tables, seal and spotting recognition, polygon layout shapes, document unwarping and image export. The markdown is the plain variant, with no HTML centering.
Validate
val() runs the shared OCR validator.
from libreyolo import LibreVLM model = LibreVLM("paddleocr-vl")metrics = model.val(data="ocr_dataset/") # det hmean, end-to-end F1, 1-NEDExport
Inference only: train() and export() raise.
Licensing
Check the license on the Hugging Face repository of the specific weights you download. Every checkpoint in the LibreYOLO org carries one, and they are not always the same across a family. That repository is the authoritative source; the summary below describes what applied when this page was last verified.
This is a description of the licenses involved, not legal advice. If the answer matters commercially, read the licenses yourself and take your own counsel.
- Original work
- PaddleOCR-VL, PaddlePaddle Authors
- Upstream license
- Apache-2.0
- Upstream source
- github.com/PaddlePaddle/PaddleOCR
- LibreYOLO code
- MIT
- Weights
- Apache-2.0, republished at huggingface.co/LibreYOLO
- Interpretation
- Apache-2.0 is a permissive license, so the PaddleOCR-VL-0.9B recognizer and the PP-DocLayoutV2 layout detector can be used in commercial and closed-source products. Redistributing them means keeping the license text and attribution notices, and the license grants a patent license. Both networks run through the Apache-2.0 Hugging Face Transformers classes; LibreYOLO ports only the document-parsing pipeline (layout post-processing, prompt routing, markdown conversion) from PaddleX at commit 4c968d30b9d0d4a693d9eaaef0881c20f8d7d0e0, also Apache-2.0 (libreyolo/models/paddleocrvl/NOTICE), and the rest of the adapter is MIT. The weights are an unmodified, ungated mirror at LibreYOLO/LibrePaddleOCRVL0.9b of PaddlePaddle/PaddleOCR-VL (revision 7fa00a8c55b735ba51ba49a9058f3f9c57a99a11) with PaddlePaddle/PP-DocLayoutV2_safetensors (revision 880e8971b88938518611c54fc0f59ad57849c9d4) under layout/, shipping LICENSE and NOTICE. The recognizer's language model is initialized from Baidu's ERNIE-4.5-0.3B, also Apache-2.0.
PaddleOCR-VL-0.9B and PP-DocLayoutV2 are Apache-2.0, code and weights, by the PaddlePaddle Authors. Commercial use is allowed. The pipeline LibreYOLO ports from PaddleX is also Apache-2.0. The weights are an unmodified, ungated mirror.
Citation
@misc{cui2025paddleocrvlboostingmultilingualdocument,
title={PaddleOCR-VL: Boosting Multilingual Document Parsing via a 0.9B Ultra-Compact Vision-Language Model},
author={Cheng Cui and Ting Sun and Suyin Liang and Tingquan Gao and Zelun Zhang and Jiaxuan Liu and Xueqing Wang and Changda Zhou and Hongen Liu and Manhui Lin and Yue Zhang and Yubo Zhang and Handong Zheng and Jing Zhang and Jun Zhang and Yi Liu and Dianhai Yu and Yanjun Ma},
year={2025},
eprint={2510.14528},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2510.14528},
}Copied from the authors' citation block at huggingface.co/PaddlePaddle/PaddleOCR-VL#citation.