NuMind builds frontier multimodal large language models (VLMs), focused on extracting information from documents and converting documents into structured JSON and clean Markdown. The product is available through a SaaS platform and also supports private enterprise deployment.
Main capabilities and uses: NuMind's models are specialized for document information extraction tasks, including named entity recognition (NER) and classification, among others. Its models lead in structured extraction and OCR, and the NuExtract series outperforms general-purpose models in benchmark tests.
Performance and deployment: NuExtract3 scores 65.2, while the best model of comparable size scores 53.8. The models have been privately deployed in document-intensive industries such as insurance, banking, and healthcare. A case study shows that AXA previously had a field-level accuracy of 30% in automated workflows, requiring extensive manual review; after using NuExtract, accuracy increased to 97% (a 23-fold reduction in errors), significantly reducing manual verification and extraction costs.
Research and releases: NuMind publishes related research, such as pretraining entity recognition encoders with LLM-labeled data, and has open-sourced OCR and structured extraction LLMs (The Reasoning Open-Source OCR & Structured Extraction LLM).
Applicable scenarios: scenarios in document-intensive industries (insurance, banking, healthcare, etc.) that require automatically extracting structured information from documents.
Limitations: The materials do not provide specific pricing models, prices, platform compatibility, usage steps, or code examples; this information is not currently available.
