skip to content
gigarouter gigarouter
models / image-to-text · coming soon

NuExtract3

numind/NuExtract3

A popular open image-to-text model, with 520.7K downloads a month. gigarouter benchmarks and hosts it as an OpenAI-compatible API.

est. price
~$1.341
/ 1k images · estimated, set at launch
API providers
0
downloads / mo
520.7K
license
apache-2.0

about this model

NuExtract3 is a 4B vision-language reasoning model for document understanding, specialized in image-to-text tasks. It combines strong structured information extraction with high-quality image-to-Markdown conversion, making it suitable for extraction pipelines, OCR, and RAG preprocessing across document types such as scans, receipts, forms, invoices, contracts, and tables.

Key strengths

  • Structured extraction from text, images, or both using a JSON template and optional instructions.
  • Document-to-Markdown conversion that preserves text, tables (HTML), math (LaTeX), and embedded images.
  • Multilingual document support.
  • Two inference modes: fast non-reasoning for deterministic extraction and reasoning mode for complex documents.
  • Template generation from natural language or input documents.

Structured extraction benchmark

Evaluated on an internal benchmark of ~600 documents (invoices, movie posters, floor plans, etc.), NuExtract3 achieved an average score of 0.651 ± 0.019 with only 27 failed outputs (non-JSON deserializable). The table below shows results compared with other models.

Structured extraction benchmark chart
ModelAverage scoreNum. failedAvg. thinking tokensAvg. answer tokens
NuExtract3 4B-RL0.651 ± 0.0192720361856
Gemma-4-E4B-it0.538 ± 0.0233130051287
Qwen3.5-9B0.479 ± 0.030170224091257
Qwen3.5-4B0.417 ± 0.031229271771201
GLM-4.6V-Flash0.435 ± 0.02615329891357
Nemotron-3-Nano-Omni0.387 ± 0.02820425827522
Ministral-3-3B0.240 ± 0.02234427586362

Document-to-Markdown evaluation

On a set of 100 documents with challenging layouts and tables, model outputs were compared against source documents using Gemini 3 Flash; rankings aligned with human votes.

Document-to-Markdown comparison chart

A two-step evaluation (convert to Markdown, then extract with Qwen3.6 27B) further confirmed that NuExtract3 preserves content and layout effectively.

Markdown-to-structured benchmark chart

NuExtract3 is hosted as a managed, OpenAI-compatible API on gigarouter, requiring no local setup — simply send an API call.

not yet live

We're benchmarking and onboarding NuExtract3 as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.