[2026/04/08] We release the OmniParsingBench, a comprehensive multimodal parsing benchmark covering Natural Image, Graphics, Audio, Natural Video, and Text-Rich Video.
[2026/02/13] 🚀🚀🚀🚀🚀 We release Logics-Parsing-v2 Model.
[2025/09/25] 🚀🚀🚀 We release Logics-Parsing Model.
Introduction
Logics-Parsing-Omni is a unified Multimodal Large Language Model (MLLM) designed to bridge the gap between pixel-level structural parsing and semantic-level cognitive captioning. It provides strong performance in both fine-grained perception and high-level semantic understanding across documents, images, audio, and video.
Logics-Parsing-Omni Overview
The construction of unified multi-modal parsing corpus and training pipeline of our proposed Logics-Parsing-Omni.
Showcase of the multifaceted capabilities of Logics-Parsing-Omni
Showcase of the multifaceted capabilities of Logics-Parsing-Omni
Key Features
Omni-Modal Unified Parsing Framework
It introduces a progressive three-level paradigm—integrating Holistic Detection, Fine-grained Recognition, and Semantic Interpretation—that fundamentally bridges the gap between pixel-based perception and logic-based cognition.
It transforms unstructured multimodal signals into a standardized, machine-readable schema that is inherently Locatable, Enumerable, and Traceable, forming an indispensable part of fact-based reasoning chains.
Surpassing traditional OCR and generic LLM pipelines, it jointly parses structural elements (e.g., dense text, layout, tables, formulas) and deep semantics (e.g., complex illustrations) with high layout fidelity.
It overcomes the bottleneck of generic image understanding by explicitly extracting dense, attribute-rich underlying data series, axis labels, and spatial topologies from scientific charts and technical diagrams to support reasoning.
Moving beyond flat ASR linear transcripts and generic video summaries, it dynamically synchronizes audio cues with critical visual contexts, explicitly capturing missed details like slides, whiteboards, and code.
It is specifically optimized for long-form educational videos, successfully extracting structured pedagogical organizations (e.g., chapter hierarchies, key concepts) and dynamic narrative logic while mitigating information redundancy and topic drift.
Powered by a meticulously constructed omni-modal dataset, the Logics-Parsing-Omni model establishes a robust balance between fine-grained structural fidelity and deep semantic interpretation.
Alongside the model, it introduces OmniParsingBench, a standardized evaluation infrastructure designed to quantitatively assess the full spectrum of parsing capabilities across documents, images, audio, and videos.
OmniParsingBench Evaluation
To rigorously evaluate the unified parsing capabilities of our model across diverse modalities, we introduce OmniParsingBench—a comprehensive, large-scale evaluation corpus. It assesses the full spectrum of parsing performance across six primary domains: Natural Image, Graphics, Document, Audio, Natural Video, and Text-Rich Video.
To provide a concise view of model capabilities, we aggregate fine-grained metrics into three core scores:
Perception (Perc.): Evaluates signal precision and structural fidelity (e.g., spatio-temporal grounding, symbol extraction).
Cognition (Cog.): Evaluates logical reasoning, semantic understanding, and hallucination resistance.
Overall (Ovr.): The comprehensive performance metric.
Full evaluation details are released in the technical report.
OmniParsingBench performance of Logics-Parsing-Omni
OmniParsingBench performance of Logics-Parsing-Omni.
Table 1: OmniParsingBench evaluation results on Image, Graphics, and Document
Model
Natural Image
Graphics
Document
Ovr.
Perc.
Cog.
Ovr.
Perc.
Cog.
Perc.
Gemini-3-Pro
61.20
55.96
66.44
87.03
84.21
87.43
87.01
GPT-5.2
39.94
37.77
42.12
82.71
69.86
91.48
77.43
Qwen3.5-397B-A17B
57.40
56.95
57.85
82.81
73.77
83.13
81.09
Qwen3-VL-235B-A22B
58.61
56.23
60.99
79.49
71.51
83.46
84.47
Qwen3-VL-30B-A3B
50.92
48.91
52.94
73.25
65.71
79.36
78.94
Qwen3-Omni-30B-A3B
47.36
46.85
47.88
77.46
70.75
78.25
73.50
Logics-Parsing-Omni(Ours)
59.07
53.77
64.37
88.66
82.01
92.12
84.90
Table 2: OmniParsingBench evaluation results on Audio and Video
Model
Audio
Natural Video
Text-Rich Video
Ovr.
Perc.
Cog.
Ovr.
Perc.
Cog.
Ovr.
Perc.
Cog.
Gemini-3-Pro
79.40
72.90
85.89
63.40
57.87
68.92
64.37
58.54
70.20
Qwen3-Omni-30B-A3B
75.17
62.13
88.22
45.23
34.15
56.32
26.86
10.22
43.50
Logics-Parsing-Omni(Ours)
79.63
69.27
89.99
61.12
56.09
66.15
69.12
57.39
80.85
Note: Bold text indicates the best result, and underlined text indicates the second-best result.
As detailed in the tables above, Logics-Parsing-Omni demonstrates highly competitive capabilities across all diverse modalities:
Dominance in Complex Modalities: Our model consistently surpasses all evaluated baselines—including the leading proprietary Gemini-3-Pro—in the Graphics, Audio, and Text-Rich Video domains.
Exceptional Cognitive Reasoning: The superiority is particularly pronounced in the Cognition metric. Logics-Parsing-Omni exhibits exceptional logical reasoning and semantic understanding, achieving top-tier scores such as 92.12 in Graphics and 80.85 in Text-Rich Video.
These results firmly validate that our model successfully bridges fundamental signal detection with complex multi-modal interpreting.
We provide a unified multimodal inference script (inference_omni.py) that supports 12 pre-defined tasks across 4 different modalities (Single Image, Multi-Image, Audio, and Video).
You can easily test different capabilities using the --task argument. Additionally, all pre-defined tasks support bilingual prompts. You can switch between English and Chinese using the --language argument (en or ch, defaults to en).
Option A: Run a Pre-defined Task
Test a specific capability using built-in prompts and assets by passing the corresponding task name and your preferred language:
shell
1# Example: Run the natural video parsing task with the English prompt (default)2python inference_omni.py --task natural_video_parsing --language en
34# Example: Run the document structure parsing task with the Chinese prompt5python inference_omni.py --task document_structure_parsing --language ch
Option B: Run a Custom Task (CLI Mode)
If you want to test your own files and prompts, use the --task custom mode along with the specific modality argument. (Note: The --language argument is ignored in custom mode since you provide the prompt directly).
shell
1# Example 1: Single Image Inference2python inference_omni.py --task custom \3 --image_paths path/to/image.jpg \4 --text_prompt "Describe the content of this image."56# Example 2: Multi-Image Inference7python inference_omni.py --task custom \8 --image_paths path/to/image1.jpg path/to/image2.jpg \9 --text_prompt "What are the differences between these two images?"1011# Example 3: Single Audio Inference12python inference_omni.py --task custom \13 --audio_path path/to/audio.wav \14 --text_prompt "Please transcribe this audio."1516# Example 4: Single Video Inference (with audio extraction)17python inference_omni.py --task custom \18 --video_path path/to/video.mp4 \19 --use_audio_in_video \20 --text_prompt "Please summarize this video."
3. Supported Pre-defined Tasks
Here is the complete list of built-in tasks you can pass to the --task argument, along with their corresponding English and Chinese prompts:
Modality
Task Argument (--task)
English Prompt (--language en)
Chinese Prompt (--language ch)
Single Image
document_structure_parsing
Output the parsing results of this document in JSON format.
以JSON格式输出此文档的解析结果。
document_structure_and_semantic_parsing
Output the parsing results of this document in JSON format. Include descriptions for illustrations, structurally parse natural images and graphics, and add a global overview at the end. Use the same language as the document text.
Please detect text and entities in the image, extract structured information such as bounding boxes, labels, attributes, and detailed descriptions, and provide a global image description. Output the results in JSON format.
Perform an in-depth parsing of the image, locate text and charts, extract their bounding boxes, labels, parsing results, and descriptions, and provide a global image description. Please present the results in JSON format.
Please detect the text and geometric shapes in the image, extract bounding boxes, labels, parsing results, and detailed descriptions, and provide a global image description. Output the results in JSON format.
Divide the audio into continuous segments primarily based on speaker and VAD (split non-speech parts by audio classification); segments should include timestamps, classification labels, ASR, and speaker IDs, with a global description added at the end, output in JSON format.
Split the video into continuous time segments based on visual semantic changes; for each segment, extract timestamps, internal audio split points and classification labels (following the principle of prioritizing human voice VAD, and classifying non-vocal parts by audio type) and video attributes. Finally, integrate a global audio-visual description, ASR (including speaker distinction), and language information. Please output in JSON format.
Describe the video content and explain its camera movement features, while simultaneously extracting the timestamps and camera movement labels of the visual segments, and output in JSON format.
描述视频内容并说明其运镜特点，同时提取视觉片段的时间戳与运镜标签，以JSON格式输出。
text_rich_video_parsing
Please analyze the video using OCR information stability as the basis for segmentation, extract the timestamp, OCR, and ASR content of each segment in chronological order, add a global audio-video description at the end, and output the result in JSON format.
Based on the input course video, generate a course description report that is clearly structured, detailed, and easy for learners to read.
根据输入的课程视频，生成一份结构清晰、内容详尽、易于学习者阅读的课程描述报告。
Multi-Image
natural_image_diff_parsing
Generate structured analysis results for the edit from the first image to the second image. List all changed elements item by item, providing corresponding bounding boxes, labels, attributes, and descriptions; finally, provide a global editing description summarizing the overall changes. Output in JSON format.
Generate the analysis results of geometric edits from the first image to the second image. The content must include structured parsing of all changed geometric elements, geometric and quantitative relationships, and provide a global editing instruction summarizing the overall changes. Output in JSON format.
