LightOnOCR-2-ocr-soup

Merged variant for extra robustness (advanced). This model combines the base checkpoint with RLVR-trained weights using task arithmetic merging, providing improved consistency across challenging document categories.

About LightOnOCR-2

LightOnOCR-2 is an efficient end-to-end 1B-parameter vision-language model for converting documents (PDFs, scans, images) into clean, naturally ordered text without relying on brittle pipelines. This second version is trained on a larger and higher-quality corpus with stronger French, arXiv, and scan coverage, improved LaTeX handling, and cleaner normalization. LightOnOCR-2 achieves state-of-the-art performance on OlmOCR-Bench while being ~9× smaller and significantly faster than competing approaches.

Highlights

• ⚡ Speed: 3.3× faster than Chandra OCR, 1.7× faster than OlmOCR, 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
• 📍 Image detection: Predicts bounding boxes for embedded images (bbox variants)

Model Variants

Benchmarks

See the paper for full benchmark details and methodology.

Usage with Transformers

Note: LightOnOCR-2 requires transformers installed from source (not yet in a stable release).

uv pip install git+https://github.com/huggingface/transformers
uv pip install pillow pypdfium2

import torch
from transformers import LightOnOcrForConditionalGeneration, LightOnOcrProcessor

device = "mps" if torch.backends.mps.is_available() else "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float32 if device == "mps" else torch.bfloat16

model = LightOnOcrForConditionalGeneration.from_pretrained("lightonai/LightOnOCR-2-1B-ocr-soup", torch_dtype=dtype).to(device)
processor = LightOnOcrProcessor.from_pretrained("lightonai/LightOnOCR-2-1B-ocr-soup")

url = "https://huggingface.co/datasets/hf-internal-testing/fixtures_ocr/resolve/main/SROIE-receipt.jpeg"

conversation = [{"role": "user", "content": [{"type": "image", "url": url}]}]

inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
)
inputs = {k: v.to(device=device, dtype=dtype) if v.is_floating_point() else v.to(device) for k, v in inputs.items()}

output_ids = model.generate(**inputs, max_new_tokens=1024)
generated_ids = output_ids[0, inputs["input_ids"].shape[1]:]
output_text = processor.decode(generated_ids, skip_special_tokens=True)
print(output_text)

Usage with vLLM

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

import base64
import requests
import pypdfium2 as pdfium
import io

ENDPOINT = "http://localhost:8000/v1/chat/completions"
MODEL = "lightonai/LightOnOCR-2-1B-ocr-soup"

# 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 1540px
• Maintain aspect ratio to preserve text geometry
• Use one image per page; batching supported by vLLM

Fine-tuning

LightOnOCR-2 is fully differentiable and supports:

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

For fine-tuning, we recommend starting with LightOnOCR-2-1B-base instead of this merged variant.

License

Apache License 2.0

Acknowlegments

The project received funding from the BPI Scribe project.

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 Cavaill\`{e}s and Baptiste Aubertin},
  year         = {2026},
  howpublished = {\url{https://arxiv.org/pdf/2601.14251}}
}