Alibaba’s Logics-Parsing-v2 Sets New OCR Benchmark Records

Model Overview

Logics-Parsing-v2, an open‑source OCR model built on the Qwen3‑VL architecture, achieves the top scores on two major benchmarks: 82.16 on the internal LogicsDocBench (900‑page PDF suite) and 93.23 on the public OmniDocBench‑v1.5. Both scores are the highest reported to date.

Parsing‑2.0 Capabilities

Flowcharts / mind maps → output in Mermaid syntax

Music scores → output in ABC notation

Code blocks / pseudocode → structured extraction

Chemical formulas → output in SMILES format

Benchmark Evaluation

LogicsDocBench consists of three scenario groups:

STEM documents – 218 pages covering physics, mathematics, engineering, etc.

Complex layout – 459 pages with multi‑column text, cross‑page tables, vertical text, and mixed graphics.

Parsing‑2.0 – 223 pages containing chemical formulas, music scores, code blocks, and flowcharts.

On this benchmark Logics‑Parsing‑v2 ranks first with an overall score of 82.16 , far ahead of competing models.

On OmniDocBench‑v1.5 the model scores 93.23 , surpassing closed‑source large models such as Gemini 2.5 Pro, GPT‑5, Qwen2.5VL‑72B, as well as specialized OCR systems like Mathpix.

Comparison with Other Models

Versus Gemini 2.5 Pro – competitive on English text (0.089 vs 0.115) and comparable on tables (0.165 vs 0.154).

Versus Mathpix – Mathpix remains stronger on formula recognition (0.06 vs 0.106), but Logics‑Parsing‑v2 shows superior overall capability.

Versus MonkeyOCR / GOT‑OCR – Logics‑Parsing‑v2 leads across all evaluated dimensions.

Versus general large models (GPT‑5, Qwen2.5VL‑72B) – dedicated OCR model demonstrates clear advantages.

The model operates end‑to‑end: an image is fed in and structured HTML is produced, eliminating the need for a multi‑stage detection‑recognition‑post‑processing pipeline.

Output Format

Results are returned as structured HTML rather than plain text. Each content block includes:

Category tags (paragraph, table, image, formula, etc.)

Pixel‑level bounding‑box coordinates

Recognized OCR text

For the new Parsing‑2.0 scenarios the output is customized:

Flowcharts → Mermaid syntax (directly renderable)

Music scores → ABC notation (readable by musicians)

Chemical formulas → SMILES format (standard chemical representation)

Deployment and Inference

conda create -n logics-parsing-v2 python=3.10
conda activate logics-parsing-v2
pip install -r requirements.txt

Download the model (choose one source):

# HuggingFace
pip install huggingface_hub
python download_model_v2.py -t huggingface

# ModelScope (faster in China)
pip install modelscope
python download_model_v2.py -t modelscope

Run inference with a single command:

python3 inference_v2.py --image_path <image_path> --output_path <output_dir> --model_path <model_path>

Demo Results

Distorted document recognition – accurate even with skewed or curved pages.

STEM document – complex formulas and charts are preserved.

Code block recognition – retains code structure.

Flowchart parsing – converts diagrams to Mermaid code ready for rendering.

Music score recognition – first OCR model to output ABC notation.

Key Takeaways

Dual‑benchmark leader: LogicsDocBench 82.16, OmniDocBench‑v1.5 93.23.

Parsing‑2.0 adds full‑cycle support for flowcharts, music scores, code blocks, and chemical formulas.

End‑to‑end single‑model pipeline: image → structured HTML without additional post‑processing.

All code and model weights are open‑source (GitHub: https://github.com/alibaba/Logics-Parsing, HuggingFace: https://huggingface.co/Logics-MLLM/Logics-Parsing-v2, ModelScope demo: https://www.modelscope.cn/studios/Alibaba-DT/Logics-Parsing/summary).