What is PP-OCRv6?
PP-OCRv6 is the sixth generation of Baidu's PaddleOCR text-recognition pipeline, released on 11 June 2026 with PaddleOCR 3.7.0 and a technical report on arXiv. It is not a vision-language model: a detector finds text lines and a separate recogniser reads each one. It comes in three tiers, tiny at 1.5 million parameters, small at 7.7 million and medium at 34.5 million, made of a 15.5M detector and a 19M recogniser.
All three share a new MetaFormer-style block. The medium recogniser pairs an LCNetV4 backbone with a light SVTR attention neck and CTC plus NRTR decoder heads. One recognition model covers Chinese, English, Japanese and Latin-script languages; the release notes say 50 languages, the model card's overview says 48.
It returns text lines with boxes and confidence scores, not Markdown. Tables, formulas and reading order need PaddleOCR's separate layout and structure models or PaddleOCR-VL.
Is PP-OCRv6 worth it?
PP-OCRv6 is the best-supported way to read plain text lines without a GPU, with the caveat that its accuracy numbers come from Baidu's in-house benchmarks. On them the medium tier scores 86.2% detection Hmean and 83.2% recognition accuracy, 4.6 and 5.1 points above PP-OCRv5_server. Baidu says the tiny tier runs 3.9 times faster than PP-OCRv5_mobile on an Intel Xeon at similar accuracy, and the release notes cite 0.13 seconds per image on an A100.
The comparison with large models needs reading carefully. Baidu's tables put GPT-5.5 at 64.2, Gemini 3.1 Pro at 71.4 and Qwen3-VL-235B at 74.9 on recognition, against 83.2 for the medium tier, and the general models score far lower on detection because they were never built to return tight text-line boxes. That shows a specialist wins at line-level OCR, not that PP-OCRv6 parses documents better than a VLM.
The only outside data point is indirect. On the independent BHL board of 2,165 historical book pages, kraken's own PP-OCRv6-medium recogniser, the same architecture trained on a different corpus, ranks 9th of 26 with a character error rate of 0.0338 and the best diplomatic score on the board, with zero looped pages. The board's maintainers say Baidu's own PP-OCRv6 will be added once its run meets their bar.
Adoption is broad. Within five months the Hugging Face hub lists community fine-tunes for Arabic, Vietnamese, Cyrillic, Hebrew and manga text, and ports to MLX, LiteRT, ExecuTorch and Rust. Next to PaddleOCR-VL-1.6 or MiniCPM-OCR 1.0 it gives up tables, formulas and Markdown, and in return runs on hardware those models cannot.
Where it earns its place
A mobile developer at a 12-person expense-tracking app
Read receipts on the phone without sending images to a server
Ship the tiny or small tier through a community LiteRT or ExecuTorch port. At 1.5M to 7.7M parameters the models fit in an app bundle, and Apache 2.0 allows commercial use without fees.
A manufacturing engineer reading serial numbers and dot-matrix codes on a production line
OCR industrial labels on CPU-only line PCs
Run the PaddleOCR pipeline with the medium detector and recogniser on CPU via OpenVINO. Baidu reports 73.3% on its industrial detection set and 77.4% on industrial recognition, so validate on your own labels before switching off manual checks.
A data engineer building a search index over millions of scanned pages
Extract raw text cheaply before sending hard pages to a VLM
Run PP-OCRv6 first and route only low-confidence or table-heavy pages to PaddleOCR-VL-1.6. Each line carries a confidence score, which gives a natural threshold for escalation.
Before you commit
- Setup
- Install PaddlePaddle, then pip install paddleocr (3.7+), and run paddleocr ocr -i image.png with the PP-OCRv6 detector and recogniser names; ONNX builds avoid the PaddlePaddle runtime.
- Integrations
- PaddleOCR CLI and Python API with optional orientation and unwarping modules; ONNX and safetensors on Hugging Face; OpenVINO on CPU; community MLX, GGUF, LiteRT, ExecuTorch and Rust ports.
- Data and privacy
- Apache 2.0 for weights and code; everything runs locally, so images never leave your hardware.
- Cost at scale
- Weights are free and tiny, so CPU servers are viable; Baidu cites 5.2 times faster CPU inference with OpenVINO, 6.1 times on an Apple M4 for the tiny tier, and 0.13 seconds on an A100.
- Limits
- Line-level output only; no layout, tables or formulas without extra PaddleOCR modules; handwriting and artistic text remain the weakest categories in Baidu's own tables.
What’s good
- 1.5M to 34.5M parameters; runs on CPUs, phones and edge boards
- Medium tier 83.2% recognition, 5.1 points above PP-OCRv5_server (in-house)
- Never loops: kraken's version scored every one of 2,165 BHL pages
- Apache 2.0, with ONNX and safetensors builds for every tier
What’s not
- Text lines and boxes only; no tables, formulas or Markdown
- All headline scores are from Baidu's in-house benchmarks
- Release notes say 50 languages, model card says 48
- Handwriting is weak: 62.1% on Chinese and 67.8% on English for the medium tier
Best for
Engineers who need fast, cheap text-line OCR on CPUs, phones or edge devices, or as the first stage of their own document pipeline.
PP-OCRv6 alternatives
PP-OCRv6: common questions
- Is PP-OCRv6 worth it?
- PP-OCRv6 reads text lines on a CPU or a phone at a fraction of a VLM's cost, as long as you only need the words and not the page structure. We rate it 4.1 out of 5.
- How much does PP-OCRv6 cost?
- Free under Apache 2.0. No first-party hosted price published.
- What are the downsides of PP-OCRv6?
- Text lines and boxes only; no tables, formulas or Markdown; all headline scores are from Baidu's in-house benchmarks; release notes say 50 languages, model card says 48; handwriting is weak: 62.1% on Chinese and 67.8% on English for the medium tier.
- Who is PP-OCRv6 best for?
- Engineers who need fast, cheap text-line OCR on CPUs, phones or edge devices, or as the first stage of their own document pipeline.
- How much does PP-OCRv6 cost for a team or in production?
- Weights are free and tiny, so CPU servers are viable; Baidu cites 5.2 times faster CPU inference with OpenVINO, 6.1 times on an Apple M4 for the tiny tier, and 0.13 seconds on an A100.
- Is PP-OCRv6 safe for company data?
- Apache 2.0 for weights and code; everything runs locally, so images never leave your hardware.
- What are the best PP-OCRv6 alternatives?
- The closest alternatives we have reviewed are Nemotron OCR v2, PaddleOCR-VL-1.6, Surya OCR 2.5 and MiniCPM-OCR 1.0.
Answers checked Nov 2, 2026.
Make PP-OCRv6? Show this rating on your site
