PaddleOCR latest version and release history

Check which version you have#

python -c "import paddleocr; print(paddleocr.__version__)"
pip index versions paddleocr          # what PyPI offers
python -m pip install -U "paddleocr==3.7.0"

paddleocr 3.7.0 requires paddlex ≥ 3.7.0 and < 3.8.0, so a fresh install today pulls PaddleX 3.7.2. The GitHub main branch moves faster than PyPI: its docs already describe HPD-Parsing (announced July 22, 2026) and newer guides, so features you read about on GitHub may not be in the pip package yet.¹²

Release cadence#

PaddleOCR 3.x releases#

1. July 22, 2026 · announcement, not a pip releaseHPD-Parsing — a lightweight vision-language model for high-throughput document parsing with hierarchical parallel decoding. Baidu reports a 4,752 tokens/s peak. It needs a customised vLLM runtime. Details.
2. June 11, 2026 · 3.7.0PP-OCRv6 becomes the default. Three tiers (tiny 1.5M / small 7.7M / medium 34.5M parameters), 50 languages in one model, and Baidu-reported gains of +4.6% detection and +5.1% recognition for medium over PP-OCRv5_server. Same day: PaddleOCR.js 0.4.2 (PP-OCRv6 in the browser) and paddleocr-mcp 0.8.5. Models.
3. May 28, 2026 · 3.6.0PaddleOCR-VL-1.6. Same architecture as 1.5, so it's a drop-in swap; Baidu reports over 96.3% on OmniDocBench v1.6. PaddleOCR-VL.
4. April 21, 2026 · 3.5.0Pluggable engines and Office files. Switch between Paddle static graph, Paddle dynamic graph and Transformers; convert Word, Excel and PowerPoint to Markdown ([doc2md]); export parsing results to DOCX; first release of the PaddleOCR.js browser SDK.
5. January 29, 2026 · 3.4.0PaddleOCR-VL-1.5. New PP-DocLayoutV3 layout model for skewed, warped and photographed pages; seal recognition and text spotting; 111 languages; cross-page table merging and heading levels.
6. October 16, 2025 · 3.3.0PaddleOCR-VL launches (0.9B: NaViT-style encoder + ERNIE-4.5-0.3B, 109 languages), plus PP-OCRv5 multilingual models for Cyrillic, Arabic, Devanagari, Telugu and Tamil.
7. August 21, 2025 · 3.2.0PP-OCRv5 English, Thai and Greek models; C++ deployment on Linux and Windows; CUDA 12 for high-performance inference; core and optional dependencies split into extras; RTX 50-series support on Windows; single-character coordinates; AI Studio and ModelScope download sources.
8. June 29, 2025 · 3.1.0PP-OCRv5 multilingual model covering 37 languages and the first MCP server.
9. May 20, 2025 · 3.0.0The big rewrite. PP-OCRv5, PP-StructureV3 and PP-ChatOCRv4; a unified inference interface built on PaddleX; PaddlePaddle 3.0. Not backward compatible with 2.x — see the migration guide.

Patch releases: 3.0.1 (June 5, 2025), 3.0.2 (June 18), 3.0.3 (June 26), 3.1.1 (Aug 15), 3.3.1 (Oct 29), 3.3.2 (Nov 13), 3.3.3 (Jan 20, 2026), 3.4.1 (Apr 14, 2026).¹

How PP-OCR evolved (2020–2026)#

All generations remain selectable in 3.7 through ocr_version="PP-OCRv3" … "PP-OCRv6", though the language coverage differs per version. PP-OCRv4 only covers ch and en.³

The 2.x line#

The last 2.x release is 2.10.0 (March 7, 2025). Its API (ocr(img, cls=True), use_gpu, PPStructure, draw_ocr) is what most older tutorials and Stack Overflow answers use. If you need it, pin it with pip install "paddleocr<3"; otherwise use our converter to move to 3.x.

FAQ#