LightOnOCR Service

A high-performance OCR microservice powered by LightOnOCR-2-1B — a state-of-the-art 1B-parameter vision-language model for converting documents (PDFs, scans, images) into clean, naturally ordered text.

Features

• State-of-the-Art OCR: Powered by LightOnOCR-2-1B, achieving top performance on OlmOCR-Bench while being ~9× smaller and significantly faster than competing approaches.
• End-to-End: Fully differentiable, no brittle OCR pipeline — the model directly converts images to text.
• Versatile: Handles tables, receipts, forms, multi-column layouts, math notation, and more.
• Dual Engine Support:
  • Hugging Face: Standard implementation for development and consumer-grade hardware.
  • vLLM: Optimized serving engine for production with high throughput (~5.71 pages/s on a single H100).
• Production Ready: Built with FastAPI, Docker, and uv for dependency management.

Deployment Configuration

The service behavior is controlled through environment variables defined in .env.

Variable	Default	Description
PROJECT_NAME	LightOnOCR-Service	Service identifier.
MODEL_SOURCE	huggingface	Selects the inference backend. Use huggingface for local dev or vllm for production.
MODEL_NAME	lightonai/LightOnOCR-2-1B	The model identifier from Hugging Face.
MODEL_CACHE_DIR	Model	Path where model weights are stored persistently.
DEVICE	cuda	Target device for inference (cuda, cpu, or mps).
MAX_FILE_SIZE_MB	10	Maximum allowed payload size for uploads.

Engine Selection Guide

Hugging Face (MODEL_SOURCE="huggingface")

• Use Case: Local development, lower VRAM availability.
• Requirements: transformers>=5.0.0, GPU with ~4GB VRAM (bfloat16) or CPU fallback.
• Characteristics: Easier to set up; uses standard PyTorch inference.

vLLM (MODEL_SOURCE="vllm")

• Use Case: Production deployment on server-grade GPUs (A10g, A100, H100).
• Setup: Requires starting a separate vLLM server:

  vllm serve lightonai/LightOnOCR-2-1B \
      --limit-mm-per-prompt '{"image": 1}' \
      --mm-processor-cache-gb 0 \
      --no-enable-prefix-caching

• Characteristics: Uses PagedAttention and optimized kernels for maximum throughput.

Installation

Local Development (with uv)

1. Clone the repository:

   git clone https://github.com/sugam24/ocr.git
   cd ocr

2. Create and activate virtual environment:

   uv venv .lighton_ocr_env
   source .lighton_ocr_env/bin/activate

3. Install dependencies:

   uv sync

4. Configure environment:

   cp .env.example .env

5. Run the server:

   uv run uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload

   The model will be downloaded automatically on first startup (~2GB).

Docker Deployment

1. Configure environment:

   cp .env.example .env

2. Build and run:

   docker-compose up --build

   The service will bind to port 8000.

API Reference

POST /api/inference

Processes an image or PDF document and extracts text using OCR.

Request

• Content-Type: multipart/form-data
• Body: file (Binary image data: JPG, PNG, PDF)

Response

{
  "text": "Extracted text content from the document...",
  "blocks": [],
  "model_version": "lightonocr-2-1b"
}

GET /info

Returns diagnostic information about the current runtime configuration.

Response

{
  "engine": "huggingface",
  "device": "cuda",
  "model": "lightonai/LightOnOCR-2-1B",
  "api_version": "v1"
}

GET /health

Health check endpoint.

Response

{
  "status": "ok"
}

Quick Test

# Upload an image for OCR
curl -X POST http://localhost:8000/api/inference \
  -F "file=@/path/to/your/document.png"

# Check service info
curl http://localhost:8000/info

Performance

• Speed: 3.3× faster than Chandra OCR, 5× faster than dots.ocr, 2× faster than PaddleOCR
• Efficiency: ~493k pages/day on a single H100 for <$0.01 per 1,000 pages
• Model Size: ~2GB (bfloat16), fits comfortably on consumer GPUs

PDF Preprocessing

PDFs are rendered at 200 DPI (scale factor ≈ 2.77) as recommended by the model authors. Aspect ratio is maintained to preserve text geometry.

Troubleshooting

Model Download Issues If the model fails to download, ensure you have internet access and sufficient disk space (~2GB). You can also manually download the model:

huggingface-cli download lightonai/LightOnOCR-2-1B --local-dir Model

Memory Issues LightOnOCR-2-1B is a 1B parameter model requiring ~2-4GB VRAM in bfloat16. If you encounter OOM errors:

• Switch to CPU: Set DEVICE="cpu" in .env
• Ensure no other processes are consuming VRAM

transformers Version LightOnOCR-2-1B requires transformers>=5.0.0. If you get import errors for LightOnOcrForConditionalGeneration, upgrade:

uv pip install --upgrade transformers

License

This service uses the LightOnOCR-2-1B model, licensed under Apache License 2.0.

Citation

@misc{lightonocr2_2026,
  title = {LightOnOCR: A 1B End-to-End Multilingual Vision-Language Model for State-of-the-Art OCR},
  author = {Said Taghadouini and Adrien Cavaillès and Baptiste Aubertin},
  year = {2026},
  howpublished = {\url{https://arxiv.org/abs/2601.14251}}
}