PaddleOCR vs Tesseract, EasyOCR, Docling and the rest

Which OCR tool for which job#

Start from the output you need, not from the tool:

Side-by-side: licenses, versions and capabilities#

The master comparison, split into three tables so each stays readable on a phone.

License and latest release#

What it is and who maintains it#

Languages and outputs#

Sources for each row are the project READMEs and PyPI pages listed at the bottom^{235671012131415}. "Not stated" means the project's own docs don't say, not that the feature is missing.

How widely each is used#

Downloads are a weak signal, but they show where the Python ecosystem has settled. Read this chart with three caveats. pytesseract is only a wrapper: Tesseract itself is installed from OS packages and isn't counted, and the wrapper is pulled into many CI jobs. RapidOCR is inflated by Docling, because a default pip install docling installs it. And all PyPI counts include CI runs and mirrors.¹⁸

The gap between rapidocr and paddleocr is partly the same models travelling through other tools: RapidOCR repackages PaddleOCR's models and Docling installs RapidOCR. PaddleOCR's own GitHub repository had 91,789 stars on the same day.²

What the benchmarks say#

There is no neutral, up-to-date benchmark that covers every tool here. The closest is OmniDocBench, published by OpenDataLab, the same organisation that maintains MinerU. Keep that in mind, and note that the dataset is licensed for research use only.¹⁶

Document parsing: OmniDocBench v1.6#

The end-to-end table scores full-page parsing: text, tables and formulas combined into one "Overall" score (higher is better). It includes PaddleOCR-VL, MinerU, Marker, olmOCR, dots.ocr, DeepSeek-OCR and the big cloud VLMs. Tesseract, EasyOCR, RapidOCR, Docling, Surya and PP-StructureV3 have no row in it, so we can't compare them here.

The top five are within 1.7 points of each other, so treat them as a group. PaddleOCR's own documents quote 96.33 for 1.6; the OmniDocBench table says 96.34. And PP-StructureV3, PaddleOCR's CPU-friendly pipeline, last appeared in the older v1.5 table with 86.73, which you can't compare directly with v1.6 scores.¹⁶

Plain text OCR: the older OmniDocBench v1.0 table#

The only shared table that includes Tesseract, EasyOCR and Surya next to PaddleOCR is OmniDocBench's component-level text recognition test on the v1.0 text subset. It is old: it evaluated PaddleOCR 2.9.1 (on PyPI since October 22, 2024), Tesseract 5.5, EasyOCR 1.7.2 and Surya 0.5.0 (August 2024).¹⁷ PaddleOCR has since moved to PP-OCRv5 and PP-OCRv6, and Surya has been rebuilt as a VLM. So use it for the broad pattern, not the exact numbers.

On English, Surya 0.5.0 had the lowest error (0.057), then PaddleOCR (0.071) and Tesseract (0.096). On Chinese and mixed-language text PaddleOCR was clearly ahead. The table also has rotated subsets and one it labels "Horizontal". There Tesseract and EasyOCR mostly failed, with edit distances of 0.92–0.997. The exception was Tesseract on 90° rotations (0.117). PaddleOCR scored 0.015 on Rotate90, 0.285 on Rotate270 and 0.021 on Horizontal.¹⁶

PaddleOCR vs EasyOCR#

These are the two most common "install it with pip and read text" libraries, and they look alike from the outside: both detect lines, recognize them and return boxes with confidence scores. The real differences:

• Activity. EasyOCR's last release is 1.7.2 from September 24, 2024, and its README says issues older than six months are closed automatically "due to limited resources". PaddleOCR shipped 3.7.0 in June 2026 with a new model generation (PP-OCRv6).⁵¹
• Framework. EasyOCR runs on PyTorch, which is convenient if your project already uses it. PaddleOCR 3.x needs PaddlePaddle or another engine. It can also run on ONNX Runtime, which in our test was 1.4–1.5× faster on CPU with identical text.²²
• Languages. EasyOCR states "80+". PaddleOCR accepts 112 language codes, and one PP-OCRv6 model covers 50 of them.²
• Accuracy. In the older OmniDocBench v1.0 text table, EasyOCR's edit distance was 0.26 on English and 0.398 on Chinese, against 0.071 and 0.055 for PaddleOCR 2.9.1. Both versions are from 2024.¹⁶
• Beyond lines. EasyOCR has no layout, table or formula features. PaddleOCR has PP-StructureV3 and PaddleOCR-VL for whole documents.

EasyOCR wins when you're already on PyTorch, need a language in its list, and want the shortest path to working code. PaddleOCR wins on maintenance, accuracy in the shared benchmark, language coverage and document features. Both are Apache-2.0.

PaddleOCR vs Tesseract#

Tesseract is the classic open-source engine, now 5.5.3 (July 24, 2026). It uses an LSTM line recognizer plus a legacy engine, reads "more than 100 languages", and its maintainers are Stefan Weil and Zdenko Podobny. Its README is candid: "in many cases, in order to get better OCR results, you'll need to improve the quality of the image".³

Where Tesseract is genuinely better is output formats. It writes plain text, hOCR, PDF, invisible-text-only PDF, TSV, ALTO and PAGE out of the box. If you are making scanned archives searchable or feeding a library system that expects ALTO, Tesseract does that in one command. It also comes as pre-built binary packages with no deep-learning framework to install.

Where PaddleOCR is better is hard input: photos, receipts, screenshots, rotated or vertical lines, and Chinese, Japanese or Korean text. In the OmniDocBench v1.0 text table, Tesseract 5.5 had an edit distance of 0.551 on Chinese versus 0.055 for PaddleOCR 2.9.1, and nearly failed on 270° rotations (0.969).¹⁶ PaddleOCR also returns structured JSON with polygons, and its document pipelines handle tables and formulas, which Tesseract does not attempt.

PaddleOCR vs RapidOCR#

The two share their models. RapidOCR's README says the project "converted the models in PaddleOCR into the highly compatible ONNX format", and that its model files "are derived from official PaddleOCR models", redistributed under Apache-2.0.⁶ So the recognition quality you get is PaddleOCR's quality. The differences are packaging and scope:

• Runtimes. RapidOCR runs on ONNX Runtime (CPU or GPU), TensorRT, OpenVINO, MNN, Paddle or PyTorch. Install with pip install rapidocr onnxruntime. That's the appeal: no PaddlePaddle wheel to match to your CPU or CUDA version.
• Scope. RapidOCR is text detection and recognition. PaddleOCR adds orientation and unwarping models, PP-StructureV3, PaddleOCR-VL, serving, an MCP server and the training code.
• Model timing. Because RapidOCR converts PaddleOCR's models, a new PaddleOCR model has to exist before RapidOCR can ship it. We didn't verify which generation RapidOCR 3.10 uses by default, so check its model list before you compare.
• Defaults. RapidOCR's default languages are Chinese and English. PaddleOCR 3.7 defaults to PP-OCRv6, which covers 50 languages in one model.

If the reason you want RapidOCR is "no PaddlePaddle", note that PaddleOCR 3.x itself accepts engine="onnxruntime". That was our fastest CPU setup for every PP-OCRv6 tier.²² RapidOCR still wins for OpenVINO, MNN or TensorRT targets, a setup with no PaddlePaddle at all, or compatibility with Docling.

PaddleOCR vs Docling#

Docling (2.143.1, MIT) is a document conversion framework rather than an OCR engine. It handles "page layout, reading order, table structure, code, formulas" and exports Markdown, HTML, lossless JSON, DocTags and more. It was started at IBM Research Zurich and is hosted by the LF AI & Data Foundation.⁷

For OCR it plugs in other engines: EasyOCR, Nemotron OCR, Tesseract, OcrMac, RapidOCR and an OnnxTR plugin. PaddleOCR isn't on that list by name. But the docling package on PyPI requires docling-slim[standard], and the standard extra includes RapidOCR. From those two files we infer that a default pip install docling runs OCR with PaddleOCR-derived ONNX models through RapidOCR.⁸ So "Docling vs PaddleOCR" is partly "Docling's pipeline vs PaddleOCR's pipeline, with similar text models underneath".

The pipelines differ in structure:

• License. Docling's code is MIT. For models, its README says to "refer to the model licenses found in the original packages". PaddleOCR is Apache-2.0 for code and every weight card we checked.
• Benchmarks. Docling has no row in OmniDocBench v1.6. Marker's README reports Docling at 50.3 on olmOCR-bench, but that is a competitor's own run.¹³ We haven't measured Docling, so we won't rank it.
• Integrations. PaddleOCR has official LangChain and MCP integrations, and its README lists Dify and RAGFlow among projects that use it.²

Choose Docling if you want one MIT-licensed converter for many document types and are happy with its default OCR. Choose PP-StructureV3 or PaddleOCR-VL if accuracy on scans, multilingual text or formulas matters most. The independent docling.org guide goes deeper on Docling itself.

PaddleOCR vs MinerU#

MinerU, from OpenDataLab, is PaddleOCR's closest competitor for PDF-to-Markdown, and the projects are entangled. MinerU's README acknowledges PaddleOCR. PaddleOCR lists MinerU among projects built on it. The PaddleOCR-VL-1.6 model card thanks MinerU and OmniDocBench, and two of PaddleOCR's own Markdown helpers are MIT code adapted "via MinerU 3.0.0".¹⁰²⁵

Accuracy and hardware#

On OmniDocBench v1.6, MinerU2.5-Pro scores 95.75 and PaddleOCR-VL-1.6 scores 96.34. MinerU-Pipeline (version 3.4.0, listed under "Pipeline Tools") scores 86.47.¹⁶ MinerU 4.0 offers four tiers. The basic tier runs ONNX models in about 2 GB of RAM ("CPU works"). The PyTorch VLM tiers need a GPU or MPS with 8 GB+ of VRAM.¹⁰ PaddleOCR-VL is GPU software too: PaddleOCR's own docs say "CPU inference is not recommended" for it.²⁵

PaddleOCR's own documentation compares PP-StructureV3 with MinerU on 925 PDF pages. On a V100, PP-StructureV3 with server models took 1.77 s per page with 17.0 GB peak VRAM, and MinerU took 1.57 s with 31.6 GB. On an A100: 1.12 s and 21.8 GB against 1.06 s and 76.9 GB. This is first-party data with old versions (PaddleOCR 3.0.0, MinerU 1.3.10), but it shows the trade-off: similar speed, much less GPU memory for PP-StructureV3.²⁴

The MinerU license history#

This is the part people get wrong. MinerU 2.5.4 (September 2025) was published on PyPI as AGPL-3.0. By 3.4.0 (June 2026) it had switched to the MinerU Open Source License, which is Apache 2.0 plus extra terms:⁹¹⁰

• a separate commercial license is required if you and your affiliates exceed 100 million monthly active users or USD 20 million monthly revenue;
• online services must "clearly and prominently indicate" that they use MinerU;
• breaking these terms ends the license.

Weights are separate again. On Hugging Face, MinerU2.5-2509-1.2B is still agpl-3.0, while MinerU2.5-Pro-2604-1.2B is apache-2.0.¹¹ For most companies the new code license is workable, but the attribution clause matters for SaaS products. PaddleOCR's Apache-2.0 terms have no revenue threshold and no display requirement.

PaddleOCR vs Surya and Marker#

Surya (0.22.1) is Datalab's OCR toolkit, now built on a 650M-parameter OCR VLM that handles layout, reading order and tables. Marker (2.0.0) is Datalab's PDF converter built on Surya. Both had new releases on July 20, 2026. Datalab reports 87.2% on "an internal benchmark set of 91 languages" and 5 pages/s on an RTX 5090. Both figures are self-reported. Surya needs vLLM on an NVIDIA GPU or llama.cpp on CPU/Apple Silicon.¹²

The license is the deciding factor for many teams. The code is Apache-2.0, but the weights use "a modified AI Pubs Open Rail-M license (free for research, personal use, and startups under $5M funding/revenue)". Above that you need a paid license from Datalab.¹² PaddleOCR's weights carry no such threshold.

On accuracy, there's no current shared number for Surya. The old v1.0 text table had Surya 0.5.0 slightly ahead of PaddleOCR 2.9.1 on English (0.057 vs 0.071) and behind on Chinese (0.123 vs 0.055); both versions are long superseded. Marker 1.8.2 scored 78.44 on OmniDocBench v1.6. On Ai2's olmOCR-Bench, published by a third competitor, Marker 1.10.1 scored 76.1 and PaddleOCR-VL 80.0, behind olmOCR v0.4.0 (82.4) and Chandra OCR (83.1). Several rows there, including PaddleOCR-VL, carry an asterisk the README doesn't explain.¹⁵¹⁶

Surya/Marker win if you are under the threshold and like their tooling and output. PaddleOCR wins if you need unrestricted commercial weights or want to stay on CPU with PP-OCRv6.

PaddleOCR vs cloud OCR APIs#

AWS Textract, Google Document AI and Azure AI Document Intelligence remove the work of running models, but charge per page and process your documents on their servers. List prices we could verify:

AWS prices are from the pricing page's worked examples for US West (Oregon). For Google, the HTML we fetched printed no unit on the OCR row, and Azure's pricing page loads prices with JavaScript, so we read Azure's from Microsoft's Retail Prices API (East US). Check both on the vendor's page before you budget.

Self-hosting isn't free either: you pay for hardware and operations. For scale, at the 0.9 s per page we measured for PP-OCRv6 small on 4 vCPUs, a million pages is about 250 hours of one such machine. Your pages and hardware will differ.²² PaddleOCR's team also runs a hosted API with a free quota of 3,000 pages per model per day, but no published price.²³ See what PaddleOCR costs and the hosted API guide.

Cloud APIs win when volume is low, you have no one to run servers, or you need their specialist extractors (invoices, IDs, queries). PaddleOCR wins on privacy, cost at volume, language coverage versus Textract, and freedom to fine-tune.

Other open models worth a look#

Four more open document models come up often. The first three are in OmniDocBench v1.6, and all four READMEs set them up on CUDA GPUs.

• dots.ocr (90.77 on v1.6): MIT code and weights, about 3B parameters.
• DeepSeek-OCR 2 (90.25): MIT code; DeepSeek-OCR-2 weights are Apache-2.0 on Hugging Face.
• olmOCR (85.74): Apache-2.0 code and weights, 7B, needs an NVIDIA GPU with at least 12 GB.
• GOT-OCR2.0 (2024): Apache-2.0 code and weights, but its README badge marks the training data CC BY-NC 4.0.

Sources: each project's README and Hugging Face license field.²⁶¹⁵¹⁶

Questions#