NuExtract3: Open-Weight 4B Vision-Language Model for Structured Extraction and OCR Under Apache 2.0

HuggingFace·medium signal

NuMind released NuExtract3, a 4B-parameter VLM based on Qwen3.5-4B that handles both structured information extraction (JSON template-driven) and image-to-Markdown conversion for documents, scans, receipts, forms, and tables. Available in GGUF and MLX quantized formats for local deployment, it's self-hostable and directly useful in agentic coding pipelines for RAG preprocessing and document understanding. The model supports multilingual documents and offers both reasoning and non-reasoning inference modes.