Evergreen Tools
|
← Back to AI Tools

LightOnOCR-3: OCR and Layout Extraction in One Model

Document understanding used to be a pipeline: one model for OCR, one for layout, another for charts. LightOnOCR-3 folds all three into one model — an empty prompt turns a page into Markdown, while the grounding prompt prepends a label and a 0–1000 normalised bounding box to every block, describes images and renders charts as HTML tables. Three sizes at 0.8B, 1B and 4B under Apache 2.0, for research and commercial use alike

Tool Interface

Interactive tool will be available soon

Features

  • ✓ Transcription and layout from one model: an empty prompt does plain transcription, the grounding prompt adds labelled boxes, image descriptions and chart data
  • ✓ Three sizes at 0.8B, 1B and 4B; the 0.8B and 4B use the Qwen3.5 vision-language architecture while the 1B keeps the previous generation's architecture
  • ✓ Bounding box coordinates are normalised to 0–1000 and blocks are tagged by type, which makes layout reconstruction straightforward
  • ✓ Layout information uses compact inline markers rather than HTML or JSON wrappers to save tokens, while chart data is still expressed as HTML tables
  • ✓ Apache 2.0 licensed, usable for research and commercial work, and self-hostable
  • ✓ Officially positioned as a low-maintenance, ready-to-use alternative to complex document understanding pipelines

How to Use

  1. Decide the task first: send an empty prompt for plain text, or the grounding prompt when you need layout, image descriptions and chart data
  2. Pick a size against your hardware: 0.8B for local or edge use, 4B when accuracy matters most, and 1B as the middle tier on the older architecture
  3. Serve it with vLLM, SGLang or Transformers; the official repository includes a minimal client, CLI and viewer
  4. When parsing output, convert the 0–1000 normalised coordinates back to pixels, then filter blocks by label such as header, footer or page_number
  5. Run a regression on your own tickets, reports and handwriting before shipping — public benchmarks are not your long tail

FAQ

How many LightOnOCR-3 models are there?

Three: LightOnOCR-3-0.8B, -1B and -4B. The 1B keeps the previous architecture while the 0.8B and 4B adopt the Qwen3.5 vision-language architecture, framed as different tradeoffs between accuracy, speed and compute. Model page: https://huggingface.co/lightonai/LightOnOCR-3-4B

What does the grounding prompt return?

It prepends labelled bounding box markers to every content block, adds image descriptions, and extracts chart values as HTML tables. Coordinates are normalised to 0–1000. Examples are in the official post at https://huggingface.co/blog/lightonai/lightonocr-3

Can it be used commercially?

Yes. It is Apache 2.0 licensed, and the official write-up says it can be used freely for both research and commercial purposes.

Does layout data cost a lot of tokens?

That is what the formatting work targets: compact inline markers for labels and coordinates instead of wrapping each region in HTML or JSON, keeping the added structure concise. Chart data is still represented as HTML tables.

How does it perform on benchmarks?

LightOn's press release says LightOnOCR-3 delivers world-leading benchmark results ahead of competing US and Chinese models, state-of-the-art performance on French-language documents and handwriting, and adds image description, chart data extraction and structure preservation over LightOnOCR-2. Release: https://www.lighton.ai/investors/documents/2026-10-08-lighton-releases-lightonocr-3-a-world-leading-model-family-for-docume.pdf — verify exact numbers against the paper and official benchmarks.