NuMind

You describe the output shape as a JSON template, and the model fills it from the document. Nested objects, arrays of line items and verbatim strings all work, which matters when a field has to match the source wording exactly. Change the template and the same document yields a different structure without retraining.

NuExtract3 was trained with reinforcement learning to think through a document before answering, and you can turn that off. Reasoning helps with merged table cells, split rows and overlapping text in messy scans. Leaving it off keeps latency and token spend down on clean, digitally native files.

The same weights handle content extraction, which means a full page becomes clean Markdown with headings, lists and tables intact. Teams use this to make legacy PDFs usable for LLM assistants and retrieval systems. It works on scans and on digital documents alike.

Templates support 20 field types, including dates, currency, country codes and booleans, so the output arrives typed rather than as loose strings. Freeform instructions and in-context examples let you steer edge cases without building a custom model.

The API charges $1 per million input tokens and $5 per million output tokens, with no onboarding delay. Volume discounts and batching bring the per-page number down for high-volume accounts. A text page runs about 1,000 tokens, while a scanned A4 page at 115 dpi lands near 1,500.

Enterprise customers can run NuExtract inside their own cloud or on-premises, with fine-tuning as an option. That matters for insurers and hospitals that cannot send documents to a third-party service. NuMind lists AXA, Crédit Mutuel, Cleareye.ai and BCA Expertise among its users.

NuExtract3 is a 4B model released under Apache 2.0, and earlier releases such as NuMarkdown-Thinking-8B are public too. You can inspect the weights, run them on your own GPUs and fine-tune without asking permission. The research behind the entity recognition model, NuNER, was published at EMNLP 2024.