NuExtract3 Turns Sensitive Docs Into Markdown Without The Cloud

NuExtract3 is a new 4-billion-parameter vision-language model that extracts structured data from documents and converts them into Markdown. It handles text, images, or both at once, making it suitable for invoices, receipts, contracts, and tables in multiple languages. The output follows JSON templates you define, so results slot neatly into any automated pipeline.

NuMind, the company behind the NuExtract series, released this open-weight model to give teams a local, private alternative for document understanding. The design focuses on keeping sensitive paperwork off cloud services while still delivering strong extraction accuracy. By running on consumer or prosumer GPUs, professionals can automate data entry for forms, scanned records, and legal documents without sending anything outside their own infrastructure.

Structured extraction and markdown conversion

Small agencies, accounting teams, and privacy-conscious professionals benefit most from this release. They can batch-process invoices, receipts, and contracts on a single GPU without exposing client data to third-party APIs. The markdown conversion also feeds clean, searchable text into retrieval systems or database indexing workflows.

Performance notes and future plans

NuMind benchmarked the 4B model against larger open and closed-source alternatives on a 600-document set that included floor plans, movie posters, and complex forms. It outperformed models several times its size, though the team cautions that enabling heavy reasoning without enough output tokens can cause smaller models to loop and fail. A public leaderboard, the evaluation dataset, and a detailed technical report are slated for release soon.

"We plan to open-source this benchmark in the coming weeks, along with a extensive leaderboard including most popular open-weight and closed-sourced APIs and a Python library allowing to easily measure model performances on structured extraction." — Source: Hugging Face