NuExtract3: 4B VLM for Document Extraction Released

NuExtract3, released by Numind, is a 4B-parameter open-weight VLM based on Qwen3.5-4B. It is optimized for extracting information from complex documents (PDFs, screenshots, forms, tables, receipts, etc.) and converting it into Markdown or JSON format.

Key features are as follows:

• Document structure understanding: Effectively handles tables, forms, and pages with complex layouts.
• Support for diverse inputs: Can process both text and visual document inputs.
• Efficient local deployment: Can run with under 4GB of VRAM, and supports various quantization formats such as GGUF, MLX, GPTQ, and FP8.
• Long-context processing: Trained for 3 days on an 8xH100 node, improving its ability to handle long documents.

It supports major inference engines such as vLLM, SGLang, and llama.cpp, and is freely available under the Apache-2.0 license.