Logics-MLLM/Logics-Parsing-v2

💻 HomePage   |   🤗 GitHub   |   🤖 Demo

Updates

• [2026/02/13] 🚀🚀🚀🚀🚀 We release Logics-Parsing-v2 Model.
• [2025/09/25] 🚀🚀🚀We release Logics-Parsing Model.

Introduction

Logics-Parsing-v2 is an advanced evolution of the previously proposed Logics-Parsing (v1). It inherits all the core capabilities of v1 model, while demonstrating more powerful capabilities on handling complex documents. Furthermore, it extends support for Parsing-2.0 scenarios, enabling structured parsing of musical sheets, flowcharts, as well as code/pseudocode blocks.

Key Features

• Effortless End-to-End Processing

  • End-to-end recognition and parsing for various kinds of document elements within a single model.
  • Handles complex-layout and text-dense documents such as newspapers and magazines with exceptional precision and ease;

• Advanced Content Recognition

  • Smaller in size, greater in performance, delivering more accurate and structured parsing of tables and scientific formulas.
  • Introducing Parsing-2.0: natively supports parsing of diverse structured content, including flowcharts, music sheets and pseudocode blocks.

• Rich, Structured HTML Output

  • Transforms documents into concise HTML -- capturing not just content, but also element types, spatial layouts, and semantic hierarchy.
  • More scientific and intuitive formats for structured elements -- such as Mermaid for flowcharts and ABC notation for musical scores.

• State-of-the-Art Performance

  • SOTA across the board: Logics-Parsing-v2 sets top records on both our in-house benchmark (overall score: 82.16) and the renowned public benchmark OmniDocBench-v1.5 (overall score: 93.23).

Benchmark

Comparisons on LogicsDocBench

We introduce LogicsDocBench, a new comprehensive evaluation benchmark comprising 900 carefully selected PDF pages, covering both traditional document Parsing-1.0 tasks and the newly introduced Parsing-2.0 scenarios. This benchmark is designed to better assess models’ capabilities in complex and diverse real-world documents parsing. The dataset is organized into three core document subsets:

• STEM Documents (218 pages):

  Focuses on high-difficulty academic and educational content, spanning over ten domains including physics, mathematics, engineering, and interdisciplinary sciences. This subset evaluates deep understanding of mathematical formulas, technical terminology, and structured knowledge representation.

• Complex Layouts (459 pages):

  Includes challenging real-world layouts such as multi-column text, cross-page tables, vertical writing, and mixed text-image arrangements. This subset comprehensively evaluate a model’s layout analysis abilities.

• Parsing-2.0 Content (223 pages):

  Targets modern digital and semi-structured content that poses significant challenges for traditional OCR systems, including:

  • Chemical Molecular formulas
  • Musical sheets
  • Code and pseudo-code block
  • Flowcharts and mind maps

For Parsing-1.0 tasks, we adopt the same evaluation protocols as OmniDocBench-v1.5 to ensure fairness and consistency across benchmarks. For Parsing-2.0, we report fine-grained results using edit distance for each subcategory, and compute an overall score as follows:

$$\text{Overall}=\frac{Parsing{1.0}^{Overall}\times 3+(1-{Chemistry}^{Edit})\times 100+(1-{Code}^{Edit})\times 100+(1-{Chart}^{Edit})\times 100+(1-{Music}^{Edit})\times 100}{7}\backslash small\; \backslash text\{Overall\}\; =\; \backslash frac\{Parsing1.0^\{Overall\}\; \backslash times\; 3\; +\; (1-\{Chemistry\}^\{Edit\})\backslash times\; 100\; +\; (1-\{Code\}^\{Edit\})\backslash times\; 100\; +\; (1-\{Chart\}^\{Edit\})\backslash times\; 100\; +\; (1-\{Music\}^\{Edit\})\backslash times\; 100\}\{7\}$$

Comprehensive evaluation of document parsing on LogicsDocBench is listed as follows:

The histogram below provides a more intuitive visualization of the advantages of our Logics-Parsing-v2 model in both Parsing-1.0 and 2.0 scenarios.

Comparisons on OmniDocBench_v1.5

We also provide the experimental results of our newly proposed Logics-Parsing-v2 model on the widely recognized open-source benchmark OmniDocBench-v1.5. As shown in the table below, Logics-Parsing-v2 achieves the highest scores among all other approaches, demonstrating its effectiveness and superiority.

* The model results in the table are sourced from the official OmniDocBench website.

Quick Start

1. Installation

conda create -n logis-parsing-v2 python=3.10
conda activate logis-parsing-v2

pip install -r requirements.txt

2. Download Model Weights

# Download our model from Modelscope.
pip install modelscope
python download_model_v2.py -t modelscope

# Download our model from huggingface.
pip install huggingface_hub
python download_model_v2.py -t huggingface

3. Inference

python3 inference_v2.py --image_path PATH_TO_INPUT_IMG --output_path PATH_TO_OUTPUT --model_path PATH_TO_MODEL

Showcases

Acknowledgments

We would like to acknowledge the following open-source projects that provided inspiration and reference for this work:

• Qwen3-VL
• OmniDocBench
• Mathpix