Infinity-Parser2 Achieves State-of-the-Art Document Parsing With New 5M-Sample Bilingual Dataset

Researchers introduce Infinity-Parser2, a multimodal AI model that combines synthetic data generation with reinforcement learning to solve document parsing at scale, releasing Infinity-Doc2-5M—a 5-million-sample bilingual corpus with comprehensive annotations. The system achieves 87.6% accuracy on olmOCR-Bench and offers two variants: Flash for low-latency inference (3.68× faster) and Pro for precision-critical applications, outperforming DeepSeek-OCR-2, PaddleOCR-VL, and MinerU2.5.

Why it matters: Enterprise document processing is a critical bottleneck in AI deployment; this breakthrough addresses the long-standing scarcity of high-quality training data while demonstrating superior performance across eight parsing tasks, positioning it as a significant advance for information extraction pipelines.