infly/Infinity-Parser2-Pro
Infinity-Parser2-Pro by infly is a flagship document understanding model optimized for maximum accuracy in precision-critical tasks. It achieves state-of-the-art results on olmOCR-Bench (87.6%) and ParseBench (74.3%), surpassing other frontier models. This model excels at robust multi-modal parsing, including document parsing, element parsing, chart parsing, chemical formula parsing, document VQA, and general multimodal understanding, leveraging an upgraded data engine and multi-task reinforcement learning.
Loading preview...
Overview
infly/Infinity-Parser2-Pro is a flagship document understanding model developed by infly, designed for high-accuracy document parsing. It is one of two variants, with the 'Pro' version specifically optimized for precision-critical tasks. The model leverages an upgraded synthetic data engine supporting both fixed-layout and flexible-layout document formats, trained on nearly 5 million diverse document parsing samples. It also incorporates Multi-Task Reinforcement Learning with a novel verifiable reward system for co-optimization of various complex tasks.
Key Capabilities
- Document Parsing: Achieves 87.6% on olmOCR-Bench and 74.3% on ParseBench, outperforming models like DeepSeek-OCR-2 and PaddleOCR-VL.
- Element Parsing: Strong performance on tasks like PubTabNet (94.76%) and UniMERNet (97.7%).
- Chart and Chemical Formula Parsing: Excels in Chart2Table (86.5%) and CoSyn_Chemical (73.19%).
- Document VQA: High accuracy on DocVQA (96.43%) and InfoVQA (86.26%).
- General Multimodal Understanding: Demonstrates robust capabilities across various multimodal benchmarks.
When to Use This Model
Infinity-Parser2-Pro is ideal for applications requiring maximum accuracy in document understanding, especially for complex and precision-critical tasks. It is suitable for parsing diverse document types, including those with intricate layouts, tables, charts, and chemical formulas. The model provides robust zero-shot capabilities across a wide range of real-world business scenarios. It primarily supports English and Chinese documents, with performance degradation for other languages or documents with multi-oriented elements.