Infinity-Parser2 surpassed DeepSeek-OCR-2 in document parsing and opened a dataset of 5 million examples

In July 2026, a technical report of Infinity-Parser2 — a multimodal model for end-to-end document parsing — appeared on arXiv. The model scored 87.6% on olmOCR-Bench and 74.3% on ParseBench, outperforming DeepSeek-OCR-2, PaddleOCR-VL-1.5, and MinerU2.5, while simultaneously releasing a corpus of 5 million annotated documents to the community.

What Infinity-Parser2 Can Do

Most OCR systems read text line by line and struggle with complex layouts — tables, mathematical expressions, diagrams, or chemical formulas. Infinity-Parser2 is trained to solve eight tasks simultaneously within a single architecture: document parsing, layout analysis, table processing, mathematical and chemical formula handling, diagram recognition, visual question answering on documents (VQA), and general multimodal understanding.

This approach allows the model to restore the complete structure of a document — not only recognizing characters but also understanding where the heading is, where the table row is, where the formula is, and in what order they should be read.

The model comes in two variants under a common architecture:

• Infinity-Parser2-Flash — optimized for low latency; throughput is 3.68 times higher than Infinity-Parser-7B
• Infinity-Parser2-Pro — maximum accuracy for mission-critical applications, including scientific and legal documentation

Benchmark results:

• olmOCR-Bench: 87.6% — best known result
• ParseBench: 74.3%
• The model outperformed DeepSeek-OCR-2, PaddleOCR-VL-1.5, and MinerU2.5

Where the Data Came From

The shortage of quality-annotated corpora is a chronic problem in document parsing. Most existing datasets cover only simple text or limited document types and rarely include scientific formulas or chemical structures. The authors of Infinity-Parser2 solved this problem by creating their own data synthesis engine: a controlled renderer generates documents of varying layouts and complexity, while an iterative refinement loop adds detailed annotations.

The result is the open corpus Infinity-Doc2-5M: 5 million examples in Chinese and English, covering scientific articles, financial reports, technical manuals, and other document types. Each sample is annotated with bounding boxes of elements, canonical forms of content (Markdown, HTML, LaTeX, SMILES, structured diagrams), and the complete reading order of pages. Such a level of detail is rarely found in open access.

How Joint Reinforcement Learning Works

The technical core of the model is a verifiable multi-task reward system. Rather than training separate specialized models for OCR, tables, and formulas, the authors apply Joint Reinforcement Learning (JRL): a single optimization function combines eight training signals and allows a single agent to learn all tasks simultaneously. The verifiability of signals means the model receives structured feedback rather than just a scalar score — this improves training stability.

"Unifying perception, structure, and reasoning in a single optimization signal" — the key architectural principle of

Infinity-Parser2.

This fundamentally sets it apart from competitors trained primarily on text recognition tasks: Infinity-Parser2 demonstrates strong results not only in OCR but also on specialized tasks — parsing chemical formulas and visual question answering on documents.

What It Means

The publication of Infinity-Parser2 along with the open corpus Infinity-Doc2-5M lowers the barrier to entry for developing document parsing systems. Teams working with PDF archives, scientific papers, financial or legal documentation now have access to a competitive baseline model and quality data for fine-tuning — without needing to build both from scratch.

Frequently Asked Questions

How does Infinity-Parser2-Flash differ from the Pro version?

Flash is optimized for speed: throughput is 3.68 times higher than Infinity-Parser-7B while maintaining acceptable quality. Pro is focused on accuracy — on it, record 87.6% on olmOCR-Bench and 74.3% on ParseBench were achieved.

What languages does the Infinity-Doc2-5M dataset support?

The corpus is bilingual: Chinese and English. It covers diverse document types with structure markup, bounding boxes, and complete page reading order.