WeVisDoc: Tencent WeChat Open-Sources End-to-End Document Parser

Tencent WeChat's vision team open-sourced WeVisDoc, an end-to-end document parser that takes a page image and outputs structured Markdown with text, LaTeX formulas, and HTML tables in a single sequence. Based on Qwen3-VL-2B/4B-Instruct, it uses a two-stage training strategy focusing on coverage then capability allocation. WeVisDoc-4B scored 95.38 on OmniDocBench and showed particular strength on degraded inputs.
Key insights
Two-Stage Training: From Coverage to Capability
WeVisDoc uses a data-centric two-stage approach. Stage I expands coverage across semantic, structural, and appearance dimensions using source-balanced heterogeneous data and structure-preserving style synthesis. A diagnostic phase identifies systematic failure clusters via residual analysis. Stage II allocates training tokens to weak areas using residual-aware weights and hard examples, without crowding out natural data.
Performance: Strong on Degraded Inputs
On OmniDocBench v1.6, WeVisDoc-4B scored 95.38, slightly ahead of HunyuanOCR-1.5 (94.74). On PureDocBench, it showed consistent gains: Clean 79.81, Digital Degraded 77.74, Real Degraded 69.08, each about 1.4 points ahead of the next best. Ablation studies show Stage II gains increase with degradation: +0.49 on Clean, +2.55 on Digital, +4.03 on Real. This confirms the residual-driven allocation strategy works best where it's needed most.