infly/Infinity-Parser-7B

VISIONPricing:Input $0.4 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:32kPublished:Oct 17, 2025Architecture:Transformer0.0K Featherless Exclusive Cold

Infinity-Parser-7B by infly is a 7 billion parameter vision-language model (VLM) specifically designed for end-to-end scanned document parsing, utilizing a context length of 32768 tokens. It employs a novel LayoutRL reinforcement learning framework with verifiable multi-aspect rewards to maintain document structure and content fidelity. This model achieves state-of-the-art performance on various document parsing benchmarks, including OCR, table and formula extraction, and reading-order detection, for both English and Chinese documents, while preserving general multimodal understanding capabilities.

Loading preview...

Infinity-Parser-7B: State-of-the-Art Scanned Document Parsing

Infinity-Parser-7B, developed by infly, is a 7 billion parameter vision-language model (VLM) engineered for high-fidelity, end-to-end scanned document parsing. It introduces a novel LayoutRL reinforcement learning framework that explicitly trains the model to be layout-aware through verifiable rewards, combining edit distance, paragraph accuracy, and reading order preservation. This approach ensures the model accurately maintains the original document's structure and content.

Key Capabilities

  • Advanced Document Parsing: Achieves new state-of-the-art performance in OCR, table and formula extraction, and reading-order detection across diverse document types and languages (English and Chinese).
  • Layout-Aware Processing: Utilizes a reinforcement learning framework to understand and preserve document layout, crucial for complex scanned documents.
  • High Fidelity: Maintains the original document's structure and content with high accuracy.
  • General Multimodal Understanding: Preserves nearly the same general multimodal understanding capability as its base model, Qwen2.5-VL-7B.
  • Extensive Training Data: Trained on the large-scale Infinity-Doc-400K dataset, which integrates synthetic data with diverse real-world samples.

Good For

  • Automated Document Processing: Ideal for tasks requiring precise extraction and understanding from scanned documents, such as invoices, reports, and academic papers.
  • Data Extraction: Excels at extracting structured data from tables and formulas within scanned documents.
  • Multilingual Document Analysis: Suitable for processing documents in both English and Chinese with high accuracy.
  • Research and Development: Provides a strong baseline for further research in document AI, particularly for models leveraging reinforcement learning and layout awareness.

Limitations

  • Currently does not provide layout or bounding box (bbox) information, limiting structured document reconstruction.
  • Lacks perception and understanding of charts and figures, preventing visual reasoning or structured extraction for graphical elements.

Finetunes
1 models