numind/NuExtract-2.0-4B

VISIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:3BQuant:BF16Context Size:32kPublished:May 26, 2025License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

NuExtract-2.0-4B by NuMind is a 3 billion parameter multimodal language model based on the Qwen2.5-VL-3B-Instruct architecture, specifically trained for structured information extraction. It excels at extracting data from both text and image inputs, supporting multilingual content and a variety of data types including strings, numbers, dates, and arrays. This model is optimized for precise data extraction using JSON templates, making it suitable for applications requiring structured output from unstructured or semi-structured sources.

Loading preview...

NuExtract 2.0 4B: Multimodal Structured Information Extraction

NuExtract 2.0 4B, developed by NuMind, is a 3 billion parameter model built upon the Qwen2.5-VL-3B-Instruct architecture. It is specifically fine-tuned for structured information extraction from diverse inputs, supporting both multimodal (text and image) and multilingual content. The model processes input text or images alongside a user-defined JSON template to extract specified data fields.

Key Capabilities

  • Structured Extraction: Extracts information into a JSON format based on a provided template.
  • Multimodal Input: Supports extraction from both text and image inputs.
  • Multilingual Support: Capable of processing content in multiple languages.
  • Flexible Template Types: Recognizes various data types including verbatim-string, string, integer, number, date-time, arrays, enum, and multi-label.
  • In-Context Learning: Improves extraction accuracy and adherence to specific formatting through in-context examples.
  • Template Generation: Can automatically generate JSON templates from XML schemas or natural language descriptions.

Good For

  • Automating data extraction from documents, forms, or images.
  • Applications requiring precise, structured output from unstructured data.
  • Developers needing a specialized model for information retrieval tasks, especially when multimodal capabilities are crucial.
  • Use cases where fine-tuning for specific extraction patterns is beneficial, with available fine-tuning tutorials.

Finetunes
3 models