0.1B-parameter OCR models that outperform multimodal LLMs

What it does

OpenOCR is a Python toolkit from Fudan University’s FVL Lab that bundles text detection, recognition, formula and table extraction, and full document parsing. It serves as both a unified training and evaluation benchmark and a collection of faithful reproductions of academic OCR papers, alongside commercial-grade systems meant for real-world use.

The interesting bit

The standout is OpenDoc-0.1B, a two-stage document parser that uses PP-DocLayoutV2 for layout analysis and the lab’s own UniRec-0.1B model to recognize text, formulas, and tables—all within a 0.1-billion-parameter budget. The README claims it scores 90.57% on OmniDocBench (v1.5), outperforming many multimodal large language models.

Key highlights

• OpenDoc-0.1B runs unified recognition of text, formulas, and tables with only 0.1B parameters
• Includes faithful core implementations from a wide range of academic papers under one benchmark
• Supports document parsing in both Chinese and English
• Functions as both a research benchmark and a deployable commercial-grade system
• Live demos hosted on HuggingFace and ModelScope

Verdict

Researchers who need reproducible paper implementations and developers who need lightweight document parsing should look here. If you want a single monolithic LLM to handle everything, this is explicitly not that.