NuExtract3

NuExtract3 is a 4-billion-parameter vision-language reasoning model for document understanding, fine-tuned from Qwen3.5-4B and licensed under Apache-2.0. It performs structured information extraction and image-to-Markdown conversion from documents such as scans, receipts, forms, invoices, contracts, and tables. The model accepts text, images, or both as input, supports multilingual documents, and offers both reasoning and non-reasoning inference modes.

Structured extraction

NuExtract3 takes an input document, a JSON template that specifies the types of fields to extract, and optional instructions. Supported field types include verbatim-string, string, integer, number, date-time, and others. The model outputs a JSON object matching the template structure; fields not found in the document return null or [].

Document-to-Markdown

NuExtract3 converts document images into clean Markdown, using HTML for tables, LaTeX for math, and <figure> tags for images.

Benchmark results

On NuMind's internal structured extraction benchmark (~600 documents including invoices, movie posters, and floor plans), NuExtract3 achieved an average score of 0.651 ± 0.019, outperforming larger models such as Qwen3.5-9B (0.479) and Gemma-4-E4B-it (0.538). The evaluation uses a tree-aligned metric: string fields are scored with indel distance, and other types with exact match. Only 27 of 600 outputs failed to parse as valid JSON.

Inference modes

NuExtract3 supports both non-reasoning (fast, deterministic) and reasoning (for complex layouts and ambiguous fields) modes. For production extraction, non-reasoning mode is recommended as the default, with reasoning enabled only for difficult examples.