PP-OCRv6 - Baidu PaddleOCR's open-source sixth-generation OCR model.

What is PP-OCRv6?

PP-OCRv6 is the sixth-generation open-source OCR model launched by the Baidu PaddleOCR team. It introduces three model sizes for the first time: Tiny (1.5M), Small (7.7M), and Medium (34.5M), covering all computing platforms from browser to server. Compared to its predecessor, detection accuracy is improved by 4.9%, and recognition accuracy by 5.1%. Specifically, Medium takes only 1.40 seconds end-to-end on an Intel Xeon CPU, 5.2 times faster than v5_server. Tiny achieves single-image prediction as low as 97ms on the browser side. The number of languages supported by a single model has expanded from 4 to 50, and new industrial scenarios such as circuit boards, digital tubes, CAD drawings, and inkjet dot matrix characters have been added.

Main functions of PP-OCRv6

• Three-tier model for flexible deploymentIt offers three sizes: Tiny (1.5M), Small (7.7M), and Medium (34.5M), which are adapted to browsers, embedded devices, and servers respectively. Choose the one that suits your needs and use it out of the box.
• High-precision text detection and recognitionIt achieves end-to-end text region localization and content recognition with a detection accuracy of 86.2% and a recognition accuracy of 83.2%, representing a generational leap forward compared to the previous generation.
• Single model coverage of 50 languagesA single model simultaneously supports Chinese, English, Japanese, and 46 Latin-based languages, allowing for the processing of multilingual mixed documents without switching models.
• Industrial scenario-specific enhancementIt adds the ability to detect and recognize complex industrial scenarios such as circuit boards, digital tubes, CAD drawings, inkjet dot matrix characters, and tire printing.
• High-speed reasoning and low latencyTiny format achieves single-image prediction as low as 97ms in a pure browser front-end; Medium takes only 1.40 seconds end-to-end on an Intel Xeon CPU, 5.2 times faster than the previous generation server model.
• High robustness and stable outputThe consistency variance of multi-size prediction detection is only 5.19%, and the consistency of edge size perturbation recognition is improved by 20.5% compared with v5. It has strong anti-interference ability to input resolution changes and cropping boundaries.
• Multiple backends and private deploymentIt supports inference backends such as OpenVINO, ONNX Runtime, and TensorRT, and can be deployed privately in intranet, offline, and security-sensitive environments.
• Structured document parsingCombined with PP-StructureV3, it can convert complex PDFs and images into Markdown or JSON formats, outputting fine-grained text coordinates and layout information.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use PP-OCRv6

• quick command line invocationAfter installing PaddleOCR 3.7.0+, execute directly. paddleocr ocr -i image.png --text_detection_model_name PP-OCRv6_medium_det --text_recognition_model_name PP-OCRv6_medium_rec This allows you to perform OCR on images.
• Python API IntegrationImport in code from paddleocr import PaddleOCRSpecify the model name during initialization, such as PaddleOCR(text_detection_model_name="PP-OCRv6_medium_det", text_recognition_model_name="PP-OCRv6_small_rec"), call predict() The method can obtain the recognition results by passing in the image path.
• Automatic model downloadWhen using a specified model for the first time, PaddleOCR will automatically download the corresponding weights from HuggingFace or ModelScope without requiring manual configuration; alternatively, you can download the model file to your local machine in advance from the ModelScope collection page or the GitHub repository.
• Browser-side frontend deploymentUsing the Tiny file (1.5M) model, inference can be performed via ONNX Runtime Web or OpenVINO front-end, enabling ultra-fast prediction of a single image in 97ms in a pure browser environment, without the need for a back-end server.
• Multi-backend inference accelerationSupports backends such as OpenVINO, ONNX Runtime, and TensorRT, allowing you to switch inference engines on demand on CPU, GPU, and embedded devices to optimize production environment performance.
• Structured document parsingCombined with the PP-StructureV3 module, after inputting images or PDFs, it can output structured results in Markdown or JSON format, including text content, coordinate positions, and page layout information.
• Select the model level as needed.Choose Tiny (mobile/browser), Small (embedded/medium computing power), or Medium (server/high precision) based on the computing power of the deployment environment. The interface of the three models is the same, and you can switch seamlessly by changing the model name.

Project address for PP-OCRv6

• Project official website:https://paddleocr.com
• ModelScope:https://modelscope.cn/collections/PaddlePaddle/PP-OCRv6

The core advantages of PP-OCRv6

• Precision leapfrog generational improvementThe text detection accuracy is improved by 4.9% compared to v5, the recognition accuracy is improved by 5.1%, the Hmean of detection in Medium mode reaches 86.2%, and the recognition accuracy reaches 83.2%, achieving a significant leap.
• Ultimate reasoning speedMedium achieves end-to-end processing time of just 1.40s on Intel Xeon CPUs, 5.2 times faster than the previous generation server model; Tiny achieves single-graph prediction time as low as 97ms on browsers, making it suitable for environments with extremely low computing power.
• Highest robustness in historyThe consistency variance of multi-size prediction detection is only 5.19% (35% lower than v5), and the consistency of edge size perturbation recognition is improved by 20.5%, with extremely strong anti-interference ability to input changes and cropping boundaries.
• Seamless coverage of the entire computing power platformThe Tiny, Small, and Medium models are available for deployment across all scenarios, from browsers and embedded devices to servers. Choose the model that best suits your needs and use it out of the box.
• Unified recognition of 50 languagesThe single model supports 50 languages, up from 4, including Chinese, English, Japanese, and 46 Latin languages, and can handle multilingual content without switching models.
• Industrial scenario deep enhancementAdded specialized recognition capabilities for real and complex industrial scenarios such as circuit boards, digital tubes, CAD drawings, inkjet dot matrix characters, and tire printing.
• Lightweight parametric crushing large modelMedium surpasses visual language models with billions of parameters, such as Qwen3-VL-235B, GPT-5.5, and Gemini-3.1-Pro, in detection and recognition accuracy with 34.5M parameters.
• Extremely low hallucination faithful reproductionThe accuracy of restoring the original text (including intentional spelling errors) is much higher than that of general large models, avoiding the "illusionary" tampering common in large models.

Comparison of PP-OCRv6 with similar competing products

Application scenarios of PP-OCRv6

• Real-time identification of mobile and edge devicesThe Tiny (1.5M) setting enables millisecond-level OCR on mobile phones, IoT devices, and embedded terminals, making it suitable for low-power scenarios such as offline scanning and express waybill recognition.
• Browser-side front-end integrationDeployed via ONNX/Web in a pure web environment, single image prediction takes as little as 97ms, enabling real-time text extraction from user-uploaded images without the need for a backend server.
• Document digitization and record managementThis tool enables batch structured extraction of documents and receipts such as ID cards, driver's licenses, business licenses, invoices, receipts, and contracts, and outputs them in JSON/Markdown format using PP-StructureV3.
• Financial and medical data entryWith an extremely low illusion rate (93.2% faithful reproduction), it accurately identifies sensitive texts such as invoices, medical records, prescriptions, and insurance policies, which have extremely high accuracy requirements.
• Industrial quality inspection and automationThe system enables automated detection and recognition of complex industrial scenarios, including circuit board silkscreen printing, digital tube reading, CAD drawing annotation, inkjet dot matrix characters, tire printing, and industrial nameplates.
• Multilingual internationalization processingThe single model covers 50 languages and is suitable for cross-border e-commerce, foreign trade documents, multilingual instructions, international content review and translation preprocessing.