Infinity-Parser2 beats DeepSeek-OCR-2 at document parsing and releases a 5-million-sample dataset

In July 2026, a technical report for 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, and the team simultaneously opened a corpus of 5 million annotated documents to the community.

What Infinity-Parser2 can do

Most OCR systems read text line by line and stumble on complex layouts such as tables, mathematical expressions, charts or chemical formulas. Infinity-Parser2 is trained to handle eight tasks at once within a single architecture: document parsing, layout analysis, table processing, math and chemical formulas, charts, visual question answering on documents (VQA) and general multimodal understanding.

This approach lets the model reconstruct a document's full structure. It does not just recognize characters; it works out which element is a heading, which is a table row, which is a formula, and in what order they should be read.

The model comes in two variants built on a shared architecture:

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

Benchmark results:

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

Where the data came from

A shortage of well-annotated corpora is a chronic problem in document parsing. Most existing datasets cover only plain text or a limited range of document types, and they rarely include scientific formulas or chemical structures. The Infinity-Parser2 authors addressed this by building their own data-synthesis engine: a controlled renderer generates documents of varying layout and complexity, and an iterative refinement loop adds detailed annotations.

The result is the open corpus Infinity-Doc2-5M: 5 million samples in Chinese and English covering scientific papers, financial reports, technical manuals and other document types. Each sample is annotated with element bounding boxes, canonical content forms (Markdown, HTML, LaTeX, SMILES, structured charts) and the full page reading order. This level of detail is rarely available in the open.

How joint reinforcement learning works

The model's technical core is a verifiable multi-task reward system. Instead of training separate specialized models for OCR, tables and formulas, the authors use Joint Reinforcement Learning (JRL): a single optimization objective combines eight training signals and lets one agent learn all tasks at the same time. Because the signals are verifiable, the model receives structured feedback rather than a single scalar score, which improves training stability.

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

Infinity-Parser2.

This sets it apart from competitors trained mainly on text recognition: Infinity-Parser2 posts strong results not only in OCR but also on specialized tasks such as chemical formula parsing and visual questions about documents.

What this means

Releasing Infinity-Parser2 together with the open Infinity-Doc2-5M corpus lowers the barrier to building document-parsing systems. Teams working with PDF archives, scientific papers, or financial and legal documentation get a competitive base model and quality fine-tuning data, without having to build both from scratch.

Frequently asked questions

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

Flash is optimized for speed: its throughput is 3.68 times higher than that of Infinity-Parser-7B, while keeping acceptable quality. Pro targets accuracy, and it is the version that reached the record 87.6% on olmOCR-Bench and 74.3% on ParseBench.

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

The corpus is bilingual: Chinese and English. It covers a variety of document types, with structure annotations, bounding boxes and the full page reading order.