Most document pipelines separate OCR, layout parsing and structured extraction; NuExtract3 collapses those steps into a single 4B vision-language model that can either output JSON following a user-provided template or generate clean Markdown suitable for downstream RAG or content workflows. That makes it easier to build end-to-end extraction services without stitching multiple specialized components.
Key Capabilities
- Template-driven structured extraction: accept a JSON template (types like verbatim-string, date, number, enums) and return a JSON matching that schema — useful for deterministic ingestion into databases or downstream logic.
- Document-to-Markdown conversion: produces Markdown for textual content, HTML for tables and LaTeX for math, enabling readable, format-preserving OCR outputs for knowledge ingestion.
- Multimodal and multilingual: handles image, text, or combined inputs across multiple languages and supports both non-reasoning (deterministic, faster) and reasoning modes (for complex layouts or ambiguous fields).
- Deployment friendly: examples and configs for vLLM, OpenAI-compatible APIs, and transformers-based local inference are provided; the model is released under Apache-2.0.
Who it's for — tradeoffs
Great fit if you need an integrated extractor for invoices, receipts, contracts, forms, tables or mixed documents and want outputs directly usable by RAG pipelines or databases. It’s also useful when you want template-first extraction (predictable JSON schemas) and an easy path to serve with vLLM or OpenAI-compatible endpoints. Look elsewhere if you need an extremely tiny on-device model, strict real-time latency on low-power devices, or a solution that requires heavy customization at the OCR-engine level (NuExtract bundles OCR+reasoning inside the model).
