NuExtract3: Document Extraction and Conversion Model

NuExtract3 is a 4B-scale unified vision-language reasoning model designed for document understanding.

Its key features are as follows:

• Structured Information Extraction: Takes text or images as input and outputs JSON format according to a JSON template.
• Markdown Conversion: Converts images and text into high-quality Markdown format.
• Multimodal and Multilingual Support: Supports combined text-and-image inputs and can process documents in various languages.
• Reasoning Mode Support: Allows you to choose between Reasoning and Non-reasoning modes.

It is optimized for building OCR, RAG preprocessing, and data extraction pipelines for various document types such as receipts, invoices, contracts, and tables. It currently supports various quantization and inference frameworks including GGUF, NVFP4, MLX, and VLLM.