Tencent's WeVisDoc-4B Tops Document Parsing Benchmarks Without OCR or Detectors
Tencent's WeChat Vision team released a compact 4B document parser that tops OmniDocBench and PureDocBench among end-to-end systems, with Apache-2.0 weights.
- Tencent released WeVisDoc-4B and 2B, end-to-end page-image-to-Markdown parsers under Apache-2.0.
- 4B model scores 95.38 on OmniDocBench v1.6 and 75.54 mean on PureDocBench, leading all four settings.
- Fine-tuned from Qwen3-VL-Instruct with no custom detector, OCR engine, or table decoder bolted on.
- Two-stage data-centric recipe: 40M-record broad coverage, then residual-diagnosed targeted refinement over ~5M records.
- Biggest gains on Real Degraded photographed pages, up 4.03 points over Stage I for the 4B model.
- Runs on vLLM 0.11.1+ with an OpenAI-compatible client; code at github.com/Tencent/WeVisDoc.
Tencent’s WeChat Vision team has released WeVisDoc-4B, an open-weights document parser that converts page images into structured Markdown, with LaTeX for equations and HTML for tables. The authors’ evaluations place the 4-billion-parameter model first among the listed end-to-end parsers on OmniDocBench v1.6 and all three PureDocBench tracks.
The accompanying project repository describes a two-stage training framework built on unmodified Qwen3-VL-Instruct 2B and 4B backbones. Inference requires no separate detector, OCR engine, table parser, or custom decoder. Both checkpoints use the Apache 2.0 license, which permits commercial deployment subject to its terms.
Top scores, including damaged pages
On the authors’ reported evaluations, where higher scores indicate better parsing, WeVisDoc-4B reaches 95.38 on OmniDocBench v1.6 and averages 75.54 across PureDocBench’s three tracks. The 2B checkpoint follows closely despite its smaller parameter count.
| Model | OmniDocBench v1.6 | PureDocBench mean |
|---|---|---|
| WeVisDoc-2B | 95.06 | 73.86 |
| WeVisDoc-4B | 95.38 | 75.54 |
PureDocBench separates clean digital pages from two harder categories: Digital Degraded includes compression artifacts and low-resolution inputs, while Real Degraded includes photographed pages, skew, and shadows. WeVisDoc-4B scores 79.81 on Clean, 77.74 on Digital Degraded, and 69.08 on Real Degraded.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
