A research team has released Infinity-Parser2, a multimodal model built to extract structured content from documents without needing large volumes of hand-labeled data.
The model's core trick is a synthetic data pipeline that generates and refines its own training examples, producing Infinity-Doc2-5M — a bilingual Chinese-English corpus of 5 million samples covering documents, tables, math formulas, chemical structures, and charts. Rather than treating each parsing task separately, Infinity-Parser2 trains across eight objectives at once using a shared reinforcement learning signal, rewarding the model for getting structure, layout, and content right simultaneously. Two variants ship: a Flash build tuned for speed (3.68x throughput improvement over the previous 7B model) and a Pro build aimed at precision-critical workflows.
Document parsing is unglamorous infrastructure work, but it underpins a surprising amount of downstream AI — RAG pipelines, financial document extraction, scientific literature ingestion. Most existing tools still struggle with tables, formulas, and mixed-language pages; Infinity-Parser2-Pro's 87.6% on olmOCR-Bench and 74.3% on ParseBench, outperforming DeepSeek-OCR-2, PaddleOCR-VL-1.5, and MinerU2.5, suggests the synthetic-data approach is closing the gap.
Benchmark leads have a habit of evaporating once real-world documents get involved, but releasing both the model and the 5-million-sample corpus under open licenses gives the research community something concrete to stress-test.