PaddleOCR-VL: Boosting Multilingual Document Parsing via a 0.9B Compact VLM (huggingface.co)

Baidu’s PaddleOCR-VL is a new, production-oriented document parsing system built around a 0.9B-parameter vision-language model that delivers state-of-the-art page- and element-level parsing while remaining resource-efficient. The core PaddleOCR-VL-0.9B couples a NaViT-style dynamic high-resolution visual encoder with a lightweight ERNIE-4.5-0.3B language model, enabling robust recognition of text, tables, formulas, charts and reading order across 109 languages. Evaluations show SOTA results on OmniDocBench v1.0/v1.5 and on multiple in-house benchmarks (OCR, table, formula, chart), where it notably outperforms traditional pipeline systems and competes with much larger VLMs — even surpassing some 72B multimodal models on chart understanding — while offering faster inference and lower compute cost. Technically, the model emphasizes efficient high-res visual features plus compact multimodal decoding to balance accuracy and throughput, making it suitable for real-world deployments (CLI/Python APIs, Docker vllm-server). The wide script coverage (Latin, Cyrillic, Arabic, Devanagari, Thai, etc.) and strong handling of degraded, handwritten, or historical documents mean teams can reduce reliance on heavy ensembles or tuning for each language/type. In short, PaddleOCR-VL promises practical, scalable document understanding for multilingual enterprise and edge scenarios where latency, cost, and broad content coverage matter.