Infinity-Parser2 Technical Report: A Breakthrough in Document Parsing Technology

Infinity-Parser2, a cutting-edge multimodal model, was unveiled in a technical report by Zuming Huang and 14 collaborators on July 8, 2026. This innovative approach aims to tackle the ongoing challenge of limited annotated parsing corpora, enhancing document parsing capabilities across various formats.

Key Contributions of Infinity-Parser2

The development of Infinity-Parser2 introduces several significant advancements. Firstly, the team has created a scalable synthesis engine that integrates a controllable rendering framework with an iterative refinement loop. This combination has led to the open-source release of Infinity-Doc2-5M, a bilingual corpus featuring 5 million samples in both Chinese and English, covering diverse document types.

Infinity-Doc2-5M is uniquely annotated with element bounding boxes, canonical content forms including Markdown, HTML, LaTeX, SMILES, and structured charts, as well as full-page reading order. This comprehensive dataset is crucial for training and improving parsing algorithms.

Innovative Reward System for Joint Reinforcement Learning

Another notable feature of Infinity-Parser2 is its multi-task reward system, which facilitates Joint Reinforcement Learning across eight co-trained objectives. These include document parsing, layout analysis, table parsing, math formula parsing, chart parsing, chemical formula parsing, document VQA, and general multimodal understanding. This integrated approach unifies perception, structure, and reasoning into a single optimization signal.

Such advancements not only improve parsing accuracy but also enhance the model's ability to understand complex document layouts and content structures.

Performance and Variants of Infinity-Parser2

Infinity-Parser2 is available in two variants: Infinity-Parser2-Flash and Infinity-Parser2-Pro. The Flash variant is optimized for low-latency inference, achieving a remarkable 3.68x throughput gain over its predecessor, Infinity-Parser-7B. In contrast, Infinity-Parser2-Pro is designed for precision-critical applications.

This latter variant has demonstrated outstanding performance, reaching a state-of-the-art accuracy of 87.6% on the olmOCR-Bench and 74.3% on ParseBench. It surpasses other models such as DeepSeek-OCR-2, PaddleOCR-VL-1.5, and MinerU2.5, showcasing robust generalization capabilities for charts, chemical formulas, and document visual question answering (VQA).