PaddleOCR-VL

PaddleOCR-VL is purpose-built as an ultra-compact vision-language model aimed squarely at the messy reality of document understanding, where pages mix printed prose, tables, equations, charts, and many languages. The design centers on a NaViT-style dynamic-resolution visual encoder paired with a lightweight language stack derived from ERNIE-4.5-0.3B, giving the roughly 0.9B parameter system an encoder-decoder shape that can ingest an image and emit structured text without ballooning into a general-purpose giant. That compact footprint is the defining design choice: the team prioritized resource efficiency so the model can run where heavier parsers cannot, while still handling the wide range of layouts a real-world document pipeline encounters, including low-resolution scans and densely packed multilingual pages spanning 109 languages. The intent is a small specialist that feels closer to a perception engine than a chat model, focused on turning pixels into clean, machine-readable structure.

The PaddlePaddle team at Baidu trained this model with progressive post-training and region-refinement techniques aimed specifically at under-optimized document regions, a recipe reflected in later checkpoints like PaddleOCR-VL-1.5 and PaddleOCR-VL-1.6 that continue to push the frontier of compact document parsing. On the OmniDocBench V1.5 benchmark the 0.9B model reaches 92.56, outperforming substantially larger systems including GPT-4o and DeepSeek-OCR on document parsing, which is a striking result given the parameter budget. In practice that combination of small size, open weights, and strong benchmark performance makes the model attractive for retrieval-augmented pipelines that need reliable extraction of tables, formulas, and multilingual text on commodity hardware. The lineage of incremental post-training releases points to a model that is expected to keep improving through targeted optimization rather than sheer scale, fitting neatly into RAG workflows and production document automation where cost, latency, and accuracy all matter.