NuExtract3

NuExtract3

Reads documents as text or images and fills a JSON template with the fields you ask for, or converts the pages to Markdown. A 4B model for teams that process scans, forms and invoices.

Made by numind on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: NuExtract3 is a 4B vision-language model for understanding documents. It does two main jobs. It fills a JSON template with facts pulled from a document, and it converts document images into Markdown. The record lists 4.5B parameters and a base of Qwen3.5-4B.

Who it suits: It suits teams that turn paper-like documents into structured data or searchable text. The card names scans, receipts, forms, invoices, contracts and tables.

Industry | Who uses it | How they use it | Example
Finance and accounting | Back-office teams | Pull fields from invoices and receipts into JSON | Filling a template with dates and totals from a scanned invoice
Legal | Contract reviewers | Extract named terms from contracts | Listing parties and clauses in a fixed JSON layout
Software | Builders of search and RAG tools | Convert scanned pages to Markdown before indexing | Turning a form scan into Markdown with tables kept

People who only need plain OCR of simple text can skip it.

Where to use it:

• Python with the transformers library on Windows, macOS or Linux
• Extraction pipelines that need JSON back
• OCR jobs on scans and photos of documents
• RAG preprocessing, where pages become Markdown

Use case | Who | What they get back
Structured extraction | Data and operations teams | JSON that follows the template you supply
Image to Markdown | Search and RAG builders | Markdown for text, HTML for tables and LaTeX for math
Template generation | Analysts | A draft template made from a description or from a document

How to set it up:

• Python on your computer
• The transformers library
• Disk space for about 9.1 GB of BF16 weights, or about 2.3 GB at 4-bit as an estimate

1. Install the Hugging Face hub package.
2. Download the model files.
3. Load the model with transformers, following the usage notes on the model card.

pip install -U huggingface_hub

hf download numind/NuExtract3

How to use it: For extraction, you give the model a document as text, images or both, a JSON template and your instructions, and it returns JSON. For conversion, you give it document images and it returns Markdown. The part of the card with run code is not in the text provided here.

Tool | What it does
Structured extraction | Fills a JSON template from text or images, guided by your instructions
Markdown conversion | Turns document images into Markdown
Reasoning mode | Lets the model think before it answers, or switches that off
Template generation | Drafts an extraction template from plain language or from a document

Example: You have a photo of a receipt and a short JSON template with fields for the shop, the date and the total. You pass both with a one-line instruction. The model returns JSON in the shape of your template.

For newcomers and experts: Think of it as a reader that fills in a form for you. A first small step is to download the files and try one clean scan with a simple template. Keep the template short at the start.

It is built on Qwen3.5-4B and supports reasoning and non-reasoning inference. On the maker's internal test of about 600 mixed documents, the card reports an average score of 0.651. It also says the reasoning of small models often fell into repetition loops, which hurt their scores.

Document as text or images
Add a JSON template and instructions
Model reads the page and reasons
JSON or Markdown comes back

Good to know:

• License: Apache 2.0.
• Cost to run: none named for the downloaded model. The card also points to a hosted API and platform, with no price given.
• Data it reaches: it downloads files from the Hugging Face Hub, then reads whatever documents you give it.
• Not for: independent proof. The benchmark is the maker's own, and the card says it plans to publish it later.
• Documents in several languages are supported.

Title: Average structured extraction score on the maker's benchmark
NuExtract3.4 4B RL: 0.651 score
gemma-4-E4B-it: 0.538 score
Qwen3.5-9B: 0.479 score
Qwen3.5-4B: 0.417 score
Ministral-3-3B: 0.240 score

On Hugging Face, 13 November 2026

• Task: image to text
• Library: transformers
• Parameters: 4.5B
• Weights: about 9.1 GB in BF16; about 2.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 40,915
• Likes: 397
• License: apache-2.0
• Based on: Qwen/Qwen3.5-4B
• Quantized versions: 18
• Fine-tunes: 12
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, vision-language, document-understanding, structured-extraction, information-extraction, document-to-markdown, markdown, reasoning, multilingual

Links

• Model: https://huggingface.co/numind/NuExtract3
• Files: https://huggingface.co/numind/NuExtract3/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download numind/NuExtract3

Then follow the usage in its model card. Nothing in this kit runs by itself.