📑 PaddleOCR-VL-1.6 Demo

Next-Generation Compact 0.9B Vision-Language Model for Document Parsing, Complex Tables, Formulas, Charts, and Text Spotting.

🏆 OmniDocBench v1.6: 96.33% SOTA ⚡ 0.9B Parameters 🎯 Region-Aware Optimization 🚀 ZeroGPU Accelerated 🤗 Model Card 🐙 GitHub 📄 Tech Report

Convert full pages, scanned papers, and reports directly into formatted Markdown with rendered LaTeX equations.

Upload Document Image

Max Tokens

256 4096

Sample Documents

Upload Document Image	Max Tokens

Target specific document elements: extract tables, LaTeX formulas, visual charts, or official stamps/seals.

Upload Element / Region

Max Tokens

256 3072

Sample Elements

Upload Element / Region	Element Task Type

Localize text bounding boxes with coordinates and view visual overlay detections.

Upload Document for Spotting

Max Tokens

512 4096

Sample Images for Spotting

Annotated Overlay

Ask questions or request custom formatting from any document image using PaddleOCR-VL-1.6.

Document Image

Max Tokens

128 2048

Sample Questions

Document Image	Prompt / Question

PaddleOCR-VL-1.6 is an upgraded compact document parsing Vision-Language Model with 0.9B parameters, built upon PaddleOCR-VL-1.5.

Region-Aware Data Optimization: Refines under-optimized regions where previous model predictions were unstable.
Progressive Post-Training: Curated data selection + reinforcement learning (RL) aligned to complex document hierarchies.
Comprehensive Multimodal Capability: Text spotting, tables, charts, mathematical formulas, seals, and ancient Chinese documents.

Model	Params	OmniDocBench v1.6	Text	Table	Formula
PaddleOCR-VL-1.6	0.9B	96.33% (SOTA)	98.4%	95.2%	95.8%
PaddleOCR-VL-1.5	0.9B	95.12%	97.2%	93.8%	94.1%
Qwen2-VL-7B-Instruct	7B	93.80%	96.5%	91.4%	92.6%
GOT-OCR-2.0	0.6B	89.20%	92.1%	85.3%	88.4%
MinerU / PDF-Extract-Kit	Pipeline	91.50%	94.2%	89.0%	90.5%