Infinity-Parser2: Multimodal Model for Advanced Document Parsing

This technical report introduces Infinity-Parser2, a sophisticated large multimodal model designed for comprehensive end-to-end document parsing. The model addresses the persistent challenge of scarce high-quality annotated parsing corpora by employing a novel, controllable data-synthesis pipeline combined with multi-task reinforcement learning. The research makes three key contributions. Firstly, it details a scalable synthesis engine that, through a controllable rendering framework and iterative refinement, created and open-sourced Infinity-Doc2-5M. This massive 5-million-sample bilingual corpus (Chinese/English) covers diverse document types, annotated with bounding boxes, canonical content forms (Markdown, HTML, LaTeX, SMILES, structured charts), and full-page reading order. Secondly, the paper presents a verifiable, multi-task reward system enabling Joint Reinforcement Learning across eight co-trained objectives, unifying perception, structure, and reasoning. Finally, two variants are released: Infinity-Parser2-Flash for low-latency inference (3.68x throughput gain) and Infinity-Parser2-Pro for precision-critical tasks, with the latter achieving state-of-the-art results on olmOCR-Bench (87.6%) and ParseBench (74.3%), surpassing competitors like DeepSeek-OCR-2 and PaddleOCR-VL-1.5, and demonstrating strong generalization across various content types.