PaddlePaddle/hpd-parsing

HPD-Parsing: Hierarchical Parallel Document Parsing

This Space demos PaddlePaddle/HPD-Parsing, a lightweight 1B-parameter document parsing model that uses a Hierarchical Parallel Decoding (HPD) paradigm to overcome the sequential bottleneck of traditional VLM-based parsers. It achieves a state-of-the-art score of 94.91% on OmniDocBench v1.6.

Upload a document image and the model will parse it into structured text (markdown), capturing layout, tables, and content hierarchy.