LightOnOCR-3

LightOn's Apache 2.0 OCR family in 4B, 1B and 0.8B sizes: Markdown transcription plus bounding boxes, image captions and chart data.

In short. LightOnOCR-3 is a family of small AI models that read scanned pages and PDFs and return the text as Markdown. In a second mode they also say where each block sits on the page, describe pictures and turn charts into tables. You pick a size, from a tiny 0.8B model to a 4B model, and the license allows commercial use.

Summary

LightOn, a French enterprise AI company, released the LightOnOCR-3 family on 8 October 2026, after LightOnOCR-2-1B in January. There are three sizes: a 4B model that LightOn recommends for most tasks, a 1B model that keeps the LightOnOCR-2 architecture and is a drop-in upgrade, and a 0.8B model for speed. The 0.8B and 4B models use the Qwen3.5 vision-language architecture, which is easier to run in common tools. The BF16 checkpoints are about 9.1 GB, 2.0 GB and 1.7 GB. All are Apache 2.0, and the card lists English as the language.

Each model has two prompt modes. With an empty prompt it transcribes the page text as Markdown, as the previous generation did. With the grounding prompt it labels every block, such as text, title, table, formula or footnote, and gives a bounding box on a 0 to 1000 page grid. It also writes a short description of each image and turns charts into HTML tables of their data. Grounding adds about 25% to the output length, which is 1,433 against 1,158 tokens per page on average for the 4B model.

LightOn's blog reports 86.3 for the 4B model on olmOCR-Bench, 1.3 points behind Infinity Parser Pro, which has 35.1B parameters, and 0.5 above Chandra 2. The 0.8B model scores 85.5 and the 1B model 84.5. On ParseBench, the 4B and 0.8B models rank first and second with 75.1 and 74.6, narrowly ahead of Infinity Parser Pro at 74.3. On a French document benchmark the 4B model leads with 74.1. LightOn ran these tests itself, and the models trail some rivals on old scans and headers or footers.

Good for

  • Converting scans and PDFs to Markdown with layout labels and bounding boxes in one model
  • Extracting data from charts and getting short descriptions of figures for search pipelines
  • Fast, small deployments with the 0.8B model, or best accuracy with the 4B model
  • Commercial use under Apache 2.0

Watch out for

  • The model card lists English only, and the blog's non-English results are limited to a French benchmark
  • The 0.8B and 4B models need a recent Transformers release, 5.5.4 or newer
  • Use only the empty prompt or the grounding prompt; other instructions are out of distribution
  • All benchmark results come from LightOn's own blog, and the 4B model scores lower than Infinity Parser Pro on olmOCR-Bench

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the LightOnOCR-3-0.8B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
LightOnOCR-3-0.8B BF16 853M 1.96 GB 0.981 GB 0.589 GB Laptop or phone
LightOnOCR-3-1B BF16 1.01B 2.32 GB 1.16 GB 0.697 GB Laptop or phone
LightOnOCR-3-4B BF16 4.54B 10 GB 5.22 GB 3.13 GB Laptop or phone

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is LightOnOCR-3?
LightOnOCR-3 is a family of small AI models that read scanned pages and PDFs and return the text as Markdown. In a second mode they also say where each block sits on the page, describe pictures and turn charts into tables. You pick a size, from a tiny 0.8B model to a 4B model, and the license allows commercial use.
Can I use LightOnOCR-3 commercially?
Yes. LightOnOCR-3 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is LightOnOCR-3 open source?
LightOnOCR-3 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run LightOnOCR-3 on my own computer?
Yes. The smallest variant, LightOnOCR-3-0.8B, needs about 0.589 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was LightOnOCR-3 released?
LightOnOCR-3 was released by LightOn on October 8, 2026.
How long a context window does LightOnOCR-3 have?
LightOnOCR-3 supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Chandra OCR 2 Datalab 5.3B 256K OpenRAIL-M (Datalab) (Conditional) 3.66 GB Laptop or phone
PaddleOCR-VL Baidu 960M – Apache 2.0 (Permissive) 0.662 GB Laptop or phone
GLM-OCR Z.ai (Zhipu AI) 900M – MIT (Permissive) 0.621 GB Laptop or phone
DeepSeek-OCR 2 DeepSeek 3.39B – Apache 2.0 (Permissive) 2.34 GB Laptop or phone