Infinity-Parser2 Technical Report

Short summary

Infinity-Parser2 is a new multimodal model for end-to-end document parsing using controllable data synthesis and multi-task RL training. The team releases Infinity-Doc2-5M, a 5-million-sample bilingual corpus with full annotations. Two model variants achieve SOTA results (87.6% on olmOCR-Bench), outperforming DeepSeek-OCR-2 and PaddleOCR.

• •New multimodal model with data-synthesis pipeline and multi-task RL for document parsing
• •Open-source 5M-sample bilingual corpus with full annotations released (Infinity-Doc2-5M)
• •Two variants (Flash for speed, Pro for accuracy) achieve SOTA: 87.6% olmOCR-Bench, 74.3% ParseBench

Generated with AI, which can make mistakes.