BizOnAI-OCR

A Korean-optimized OCR model for industrial document processing developed by ONTHEIT.

Built on Qwen3-VL-8B, BizOnAI-OCR is tuned for real-world Korean documents — contracts, medical records, financial forms, government paperwork, and more. It handles Korean-specific layouts, mixed Korean/English/Chinese text, decorative spacing, vertical tables, and complex forms that trip up general-purpose OCR models.

Highlights

• Korean-first: fine-tuned on Korean industrial documents
• Bilingual: remains competitive on English OCR benchmarks
• Markdown output: structured output with tables, headings, and formatting preserved
• vLLM-ready: served efficiently via vLLM (OpenAI-compatible API) or plain transformers

Benchmark Results

KDoc-OCRBench (Korean)

Evaluated on KDoc-OCRBench — 14,738 unit-test-style assertions across 804 Korean industrial PDFs.

Model	Baseline	Header/Footer	Long Text	Table	Overall
BizOnAI-OCR	98.1	88.8	77.0	68.0	83.0
olmOCR v0.2.0	99.9	88.5	73.7	53.4	78.9
PaddleOCR-VL	100.0	90.5	65.7	57.6	78.4
DeepSeek OCR	99.8	88.3	66.0	53.1	76.8
GLM OCR	99.9	96.3	41.8	31.0	67.2

olmOCR-bench (English)

Evaluated on olmOCR-bench by Allen AI.

Model	ArXiv	Old Scans Math	Tables	Old Scans	Headers & Footers	Multi Column	Long Tiny Text	Base	Overall
BizOnAI-OCR	84.0	74.2	89.7	47.3	92.4	80.2	92.5	98.7	82.4
PaddleOCR-VL	85.7	71.0	84.1	37.8	97.0	79.9	85.7	98.5	80.0
olmOCR v0.2.0	78.8	77.5	71.9	45.4	94.2	78.6	81.4	99.8	78.5
DeepSeek OCR	77.2	73.6	80.2	33.3	96.1	66.4	79.4	99.8	75.7

Quick Start

Full installation instructions, CLI tools, and the benchmark pipeline are in the GitHub repository.

Serve with vLLM (recommended)

python -m vllm.entrypoints.openai.api_server \
    --model ONTHEIT/BizOnAI-OCR \
    --port 8000 \
    --dtype bfloat16 \
    --max-model-len 16384 \
    --trust-remote-code \
    --served-model-name bizonai-ocr

Python CLI

git clone https://github.com/ONTHEIT-AI/BizOnAI-OCR.git
cd BizOnAI-OCR && uv venv && source .venv/bin/activate
uv pip install -e .
bizonai-ocr input.pdf ./output

Tested Environment

Python 3.12, vllm==0.11.0, transformers==4.57.6, torch==2.8.0+cu128 (CUDA 12.8)

Credits

• Qwen3-VL by Alibaba
• olmocr by Allen AI
• olmOCR-bench by Allen AI
• vLLM

License

This repository is released under the Apache2.0 and OpenRAIL. See the LICENSE file for details.