AI Briefing
KO

Numind Unveils 4B VLM 'NuExtract3'

·2026.05.22 19:07

Key point

Numind has released NuExtract3, a 4B-scale open-weight VLM specialized for Markdown conversion and structured data extraction from documents.

1 / 2

Details

Numind has released NuExtract3, built on Qwen3.5-4B, under the Apache-2.0 license. This model is a 4B-scale vision language model (VLM) optimized for extracting information from documents with complex visual structures, such as PDFs, screenshots, forms, and receipts.

Key features and characteristics are as follows:

  • Markdown conversion of document images and structured data extraction
  • Support for data extraction according to user-specified JSON templates
  • Ability to process tables, forms, and layout-heavy pages
  • Enhanced long-context handling through training on 8xH100 GPUs for 3 days

NuExtract3 is the successor to the existing NuMarkdown, offering developers a useful open-weight alternative for building document extraction pipelines that can run in local environments.

ocrvlmmarkdownopen-weightsdocument-extractionnumind
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

InflearnInflearn Develops 'Video Screen Translation' Feature That Translates Text Within Video FramesPyTorchKRXberg, 101-format extraction enginePyTorchKRTrillion Labs Releases GravityOCR, an OCR Model Applying Self-Speculative DecodingHacker NewsApple Releases LensVLM-9B for Processing Compressed Text ImagesHuggingFace BlogLiquid AI Releases LFM2.5-VL-DSpark: An Experimental Draft Model for Faster Vision-Language InferenceTLDR AIInvolution and Overseas Expansion of Chinese AI: Structural Limitations and US Response Strategies