PP-OCRv6 on Hugging Face: 50-Language OCR from 1.5M to 34.5M Parameters

PP-OCRv6 is the latest generation of PaddleOCR’s universal OCR model family, now available on Hugging Face. It’s designed for real-world text detection and recognition across documents, screenshots, multilingual images, digital displays, industrial labels, and scene text.

The model family scales from 1.5M to 34.5M parameters, with three tiers: tiny, small, and medium. The medium and small tiers support 50 languages, including Simplified Chinese, Traditional Chinese, English, Japanese, and 46 Latin-script languages. You can try PP-OCRv6 online quickly: PP-OCRv6 Online Demo.

On PaddleOCR’s official in-house multi-scenario OCR benchmarks, PP-OCRv6_medium reaches 86.2% detection Hmean and 83.2% recognition accuracy. Compared with PP-OCRv5_server, it improves text detection by +4.6 percentage points and text recognition by +5.1 percentage points.

PP-OCRv6 focuses on a practical OCR need: producing accurate, structured text outputs with small models and flexible deployment options. For a deeper discussion of why specialized OCR models remain useful in the VLM era, see our previous blog: PP-OCRv5 on Hugging Face: A Specialized Approach to OCR.

What’s new in PP-OCRv6

PP-OCRv6 introduces architecture, training, and data improvements across detection and recognition. The main design goal is to improve OCR accuracy while keeping model sizes suitable for different deployment settings.

Three model tiers

PP-OCRv6 provides three model tiers, covering different model sizes and OCR accuracy levels.

| Model | Model size | Detection Hmean | Recognition accuracy | Typical application scenarios |
|——-|————|—————–|———————-|——————————-|
| PP-OCRv6_tiny | 1.5M params | 80.6% | 73.5% | Edge devices, lightweight local OCR, latency-sensitive demos, constrained environments |
| PP-OCRv6_small | 7.7M params | 84.1% | 81.3% | Mobile, desktop, balanced OCR services, multilingual OCR with lower compute cost |
| PP-OCRv6_medium | 34.5M params | 86.2% | 83.2% | Accuracy-oriented OCR, server-side pipelines, industrial OCR, document ingestion, multilingual OCR |

PPLCNetV4 backbone

PP-OCRv6 uses PPLCNetV4 as a unified backbone for text detection and text recognition.

For developers, the main benefit is consistency across the model family. The tiny, small, and medium tiers are not unrelated models; they are part of the same OCR family and share a common architectural direction.

RepLKFPN for text detection

Text detection is the first stage of the OCR pipeline. Detection quality affects the crops sent to the recognizer, and poor crops often lead to poorer recognition.

PP-OCRv6 upgrades the detection module with RepLKFPN, a lightweight large-kernel feature pyramid network designed for multi-scale text detection while keeping inference efficient.

This is relevant for real-world OCR inputs, where text may be small, dense, rotated, low-resolution, or embedded in complex backgrounds.

EncoderWithLightSVTR for recognition

For text recognition, PP-OCRv6 uses EncoderWithLightSVTR. It combines local context modeling with global attention to improve recognition quality on challenging text crops.

The recognition improvements are especially relevant for multilingual text, screen text, industrial characters, special symbols, dense text, and noisy image regions.

Unified multilingual OCR

The medium and small tiers support 50 languages in one model family, covering Simplified Chinese, Traditional Chinese, English, Japanese, and 46 Latin-script languages.

This helps reduce the need for separate OCR models across common multilingual OCR scenarios.

Quick start with PaddleOCR

Install PaddleOCR:

pip install paddleocr

Run OCR with Paddle Inference (Default backend):

from paddleocr import PaddleOCR

ocr = PaddleOCR(
    use_doc_orientation_classify=False,
    use_doc_unwarping=False,
    use_textline_orientation=False,
)
result = ocr.predict("https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_002.png")

for res in result:
    res.print()
    res.save_to_img("output")
    res.save_to_json("output")

The OCR result can be saved as visualization images and structured JSON output. The structured output can then be used by downstream systems such as document parsing, search, extraction, RAG, analytics, or agent workflows.

Available inference backends

PP-OCRv6 can be used with multiple inference backends through PaddleOCR. PaddleOCR 3.7 provides a unified inference-engine interface, where engine selects the underlying runtime and related configuration can be passed through the pipeline or module API.

For Hugging Face users, PaddleOCR supports running selected OCR and document parsing models with a Transformers backend. This can be enabled with:

engine="transformers"

For more details on how the Transformers backend works in PaddleOCR, see:

PaddleOCR: Running OCR and Document Parsing Tasks with a Transformers Backend

Run PP-OCRv6 example with Transformer Backend:

from paddleocr import PaddleOCR

ocr = PaddleOCR(engine="transformers")
result = ocr.predict("https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_002.png")

for res in result:
    res.print()
    res.save_to_img("output")
    res.save_to_json("output")

Why this matters

PP-OCRv6 is a solid step forward for practical OCR. The 50-language support in a single model family is a big deal for anyone dealing with multilingual documents. The tiered approach means you’re not forced to use a heavyweight model for simple tasks, and the tiny model at 1.5M parameters is genuinely useful for edge devices.

The detection improvements (+4.6 points over v5) and recognition gains (+5.1 points) are meaningful, especially for challenging inputs like dense text or noisy images. The RepLKFPN and EncoderWithLightSVTR changes aren’t just marketing fluff; they address real pain points in OCR pipelines.

If you’ve been using separate models for different languages or struggling with accuracy on complex layouts, PP-OCRv6 is worth a look. The Hugging Face integration makes it easy to test without setting up a full PaddlePaddle environment, and the ONNX Runtime backend gives you deployment flexibility without vendor lock-in.

Give it a try and see if it handles your use case better than the alternatives.