Skip to main content
Back

LightOn Releases LightOnOCR-3 to Parse Documents Without Extra Tools

LightOn released a three-size OCR family that transcribes text, detects layout with bounding boxes, describes images and extracts chart data in one pass.

Model, Full Access
  • LightOn released LightOnOCR-3 in 0.8B, 1B and 4B sizes, all under Apache 2.0.
  • New grounding mode returns labeled bounding boxes, image descriptions and chart data as HTML tables.
  • 0.8B and 4B use Qwen3.5 VL architecture; 1B keeps LightOnOCR-2 architecture as a drop-in upgrade.
  • Claims leading open-weight scores on OlmOCR-Bench and ParseBench, though third-party validation is still pending.
  • Trained with MinHash dedup, oversampled tables and handwriting, synthetic DocDuck data, Muon optimizer and multi-objective RLVR.
  • Runs on vLLM and SGLang; render PDFs at 400 DPI, longest edge 2048 px.

LightOn has released LightOnOCR-3, a family of compact vision-language models for document parsing. The models transcribe pages, locate detected regions, describe images, and convert chart values into tables. That combination can consolidate several stages of a document pipeline while preserving coordinates that link extracted content to its source on the page.

Three checkpoints, two architectures

All three checkpoints are available on Hugging Face under the Apache 2.0 license. The 1B model retains the previous architecture and targets compatibility with LightOnOCR-2 deployments. The 0.8B checkpoint and 4B model use the Qwen3.5 vision-language architecture.

Checkpoint Architecture Intended tradeoff
0.8B Qwen3.5 VLM Lowest compute and memory requirements
1B Previous LightOnOCR architecture Drop-in upgrade path from LightOnOCR-2
4B Qwen3.5 VLM Higher capacity at greater serving cost

A prompt switch adds page geometry

The models expose two supported prompt configurations:

Mode Text prompt Output
Transcription Empty Page text
Grounding grounding Text, region labels, bounding boxes, image descriptions, and chart tables

Grounding mode prepends each detected block with a type label and bounding box in the form (x1,y1,x2,y2). Coordinates are normalized to a 0-to-1000 range, independent of the source image dimensions. Applications can convert them back to pixels by multiplying the horizontal values by the page width and the vertical values by the page height, then dividing by 1000.