AI · AI model
WeVisDoc
Open-source vision-language model converting complex document images to structured Markdown
WeVisDoc is an end-to-end vision-language model developed by Tencent's WeChat Vision Team for document parsing and OCR. Built on Qwen-VL architecture in 2B and 4B sizes, it transforms scanned page images into structured Markdown complete with LaTeX formulas and HTML tables. It is designed to maintain high parsing accuracy even on blurry, distorted, or heavily degraded documents.
What it's for
- Convert scanned pages and PDFs into clean Markdown documents
- Extract mathematical formulas directly into LaTeX formatting
- Parse complex, multi-row financial tables into structured HTML
- Transcribe low-quality or degraded real-world document scans
Features
- End-to-end image-to-Markdown document parsing
- LaTeX formula extraction and HTML table formatting
- Available in 2B and 4B parameter sizes based on Qwen-VL
- High robustness on degraded and distorted scans
- vLLM server support with OpenAI-compatible API endpoint
- Local inference script support via Hugging Face Transformers
- Support for PNG, JPEG, and WebP input formats
Pricing
Open source
Details
- Type
- AI model
- Category
- AI
- Topic
- Document Parsing
- Open source
- Yes
- Platforms
- Linux, Self-hosted, API, Command line
- For
- Developers and ML engineers building document extraction pipelines and RAG workflows.
- Alternative to
- DeepSeek-OCR, olmOCR, FireRed-OCR, Nougat
- Tags
- ocr, document parsing, vlm, vision language, latex, markdown, open source
Why it was saved
- @isaac_flath: Solves common OCR failures on difficult scans like financial statements and mathematical formulas by using targeted data augmentation during training.
Related
Open-source vision-language model converting complex documents into structured Markdown using reasoning.
Open 10B multimodal model with parallel reasoning competitive with frontier systems
Open-source vision-language model for parsing multi-page documents in one pass
Agentic document parsing and extraction platform for AI pipelines.
Browser that installs and runs open-source AI apps with one click
Open-weights fine-tune of Llama 3.1 built for agentic function calling.
Updated