You can deploy this model and start earning money today!

lightonai/
LightOnOCR-0.9B-32k-1025

Name: LightOnOCR-0.9B-32k-1025
Rating: 2.5 (18 reviews)

document-understanding

Views

No views yet

[!NOTE] 🚀 LightOnOCR-2 is now available and state-of-the-art on OlmOCR-bench, with new image detection variants! Check it out here: lightonai/LightOnOCR-2-1B

LightOnOCR-0.9B-32k-1025

Pruned-vocabulary version (32k tokens) optimized for European languages, offering additional speedup with minimal accuracy loss.

LightOnOCR-1B is a compact, end-to-end vision–language model for Optical Character Recognition (OCR) and document understanding. It achieves state-of-the-art accuracy in its weight class while being several times faster and cheaper than larger general-purpose VLMs.

📄 Paper | 📝 Read the full blog post | 🚀 Try the demo

Highlights

⚡ Speed: 5× faster than dots.ocr, 2× faster than PaddleOCR-VL-0.9B, 1.73× faster than DeepSeekOCR
💸 Efficiency: Processes 5.71 pages/s on a single H100 (~493k pages/day) for <$0.01 per 1,000 pages
🧠 End-to-End: Fully differentiable, no external OCR pipeline
🧾 Versatile: Handles tables, receipts, forms, multi-column layouts, and math notation
🌍 Compact variants: 32k and 16k vocab options for European languages

Model Overview

LightOnOCR combines a Vision Transformer encoder(Pixtral-based) with a lightweight text decoder(Qwen3-based) distilled from high-quality open VLMs. It is optimized for document parsing tasks, producing accurate, layout-aware text extraction from high-resolution pages.

Benchmarks

Model	ArXiv	Old Scans	Math	Tables	Multi-Column	Tiny Text	Base	Overall
LightOnOCR-1B-1025 (151k vocab)	81.4	71.6	76.4	35.2	80.0	88.7	99.5	76.1
LightOnOCR-1B-32k (32k vocab)	80.6	66.2	73.5	33.5	71.2	87.6	99.5	73.1
LightOnOCR-1B-16k (16k vocab)	82.3	72.9	75.3	33.5	78.6	85.1	99.8	75.4

All benchmarks evaluated using vLLM.

Installation

[2025/11/24] 🚀 LightOnOCR is now officially supported in vLLM v0.11.1 🚀

bash

uv venv --python 3.12 --seed
source .venv/bin/activate

# install any version higher than 0.11.1
uv pip install vllm==0.11.2
# extra deps need only to run the example below
uv pip install pypdfium2 pillow requests

Start Server

bash

vllm serve lightonai/LightOnOCR-0.9B-16k-1025 \
    --limit-mm-per-prompt '{"image": 1}'

PDF Inference

python

import base64
import requests
import pypdfium2 as pdfium
import io

ENDPOINT = "http://localhost:8000/v1/chat/completions"
MODEL = "lightonai/LightOnOCR-0.9B-16k-1025"

# Download PDF from arXiv
pdf_url = "https://arxiv.org/pdf/2412.13663"
pdf_data = requests.get(pdf_url).content

# Open PDF and convert first page to image
pdf = pdfium.PdfDocument(pdf_data)
page = pdf[0]
# Render at 200 DPI (scale factor = 200/72 ≈ 2.77)
pil_image = page.render(scale=2.77).to_pil()

# Convert to base64
buffer = io.BytesIO()
pil_image.save(buffer, format="PNG")
image_base64 = base64.b64encode(buffer.getvalue()).decode('utf-8')

# Make request
payload = {
    "model": MODEL,
    "messages": [{
        "role": "user",
        "content": [{
            "type": "image_url",
            "image_url": {"url": f"data:image/png;base64,{image_base64}"}
        }]
    }],
    "max_tokens": 4096,
    "temperature": 0.2,
    "top_p": 0.9,
}

response = requests.post(ENDPOINT, json=payload)
text = response.json()['choices'][0]['message']['content']
print(text)

Rendering and Preprocessing Tips

Render PDFs to PNG or JPEG at a target longest dimension of 1280–1300 px
Maintain aspect ratio to preserve text geometry
LightOnOCR is robust to moderate skew; heavy rotation correction is optional
Use one image per page; batching supported by vLLM

Variants

Variant	Description
LightOnOCR-1B-1025	Full multilingual model (default)
LightOnOCR-1B-32k	Fastest pruned-vocabulary version (32k tokens) optimized for European languages
LightOnOCR-1B-16k	Most compact variant with smallest vocabulary

Fine-tuning

Transformers integration is coming soon for training.

LightOnOCR is fully differentiable and supports:

LoRA
Domain adaptation (receipts, scientific articles, forms, etc.)
Multilingual fine-tuning with task-specific corpora

Example fine-tuning configurations will be released alongside the dataset.

Data

Trained on a diverse large-scale PDF corpus covering:

Scientific papers, books, receipts, invoices, tables, forms, and handwritten text
Multiple languages (Latin alphabet dominant)
Real and synthetic document scans

The dataset will be released under an open license.

License

Apache License 2.0

Acknowlegments

The project received funding from the BPI Scribe project.

Citation

@misc{lightonocr2025,
  title        = {LightOnOCR-1B: End-to-End and Efficient Domain-Specific Vision-Language Models for OCR},
  author       = {Said Taghadouini and Baptiste Aubertin and Adrien Cavaillès},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/blog/lightonai/lightonocr}}
}

lightonai/LightOnOCR-0.9B-32k-1025