PP-OCRv6 scales from 1.5M to 34.5M parameters across 50 languages
PaddlePaddle's PP-OCRv6 adds 50-language support across four model sizes, with the smallest checkpoint running at 1.5M parameters and the largest at 34.5M, all now available on Hugging Face.
PaddlePaddle has released PP-OCRv6, a text recognition pipeline that now covers 50 languages in a single release, with model sizes ranging from 1.5 million to 34.5 million parameters depending on accuracy and speed requirements.
PP-OCRv6 handles detection, recognition, and layout analysis in one workflow. The four checkpoints—PP-OCRv6-mobile (1.5M params), PP-OCRv6-server (6.9M), PP-OCRv6-large (17.2M), and PP-OCRv6-xlarge (34.5M)—let practitioners trade inference speed for recognition accuracy. All four support the same 50 languages, including English, Chinese, Japanese, Korean, Arabic, Cyrillic scripts, and Latin-alphabet European languages. The model cards on Hugging Face list parameter counts, supported languages, and recommended use cases for each size.
The mobile checkpoint targets edge devices and real-time applications where sub-10ms latency matters. The server and large variants sit in the middle for cloud deployments that need higher accuracy without maxing out GPU memory. The xlarge checkpoint is the accuracy ceiling, designed for batch processing where throughput matters more than single-image speed. The v6 series improves multi-script handling over v5, particularly for documents that mix Latin and non-Latin characters on the same page.
The weights are Apache 2.0 licensed and ship with pre-built inference scripts for Python. The Hugging Face repository includes sample images, a quick-start notebook, and links to the full PaddleOCR toolkit for custom training. Practitioners running local OCR pipelines now have a 50-language option that scales from Raspberry Pi to datacenter without switching architectures.
