LightOnOCR 1B 1025

Full BF16 version of the model. We recommend this variant for inference and further fine-tuning.

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](https://huggingface.co/lightonai/LightOnOCR-1B-1025/blob/main/ https://arxiv.org/pdf/2601.14251) | 📝 Read the full blog post | 🚀 Try the demo | 📓 Finetuning notebook

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

All benchmarks evaluated using vLLM on the Olmo-Bench.

Installation

VLLM

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

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


uv pip install vllm==0.11.2

uv pip install pypdfium2 pillow requests

Start Server

vllm serve lightonai/LightOnOCR-1B-1025 \
    --limit-mm-per-prompt '{"image": 1}' --mm-processor-cache-gb 0 --no-enable-prefix-caching

PDF Inference

import base64
import requests
import pypdfium2 as pdfium
import io

ENDPOINT = "http://localhost:8000/v1/chat/completions"
MODEL = "lightonai/LightOnOCR-1B-1025"


pdf_url = "https://arxiv.org/pdf/2412.13663"
pdf_data = requests.get(pdf_url).content


pdf = pdfium.PdfDocument(pdf_data)
page = pdf[0]

pil_image = page.render(scale=2.77).to_pil()


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


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)

Transformers

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

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

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)

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

Variants

Fine-tuning

Transformers integration is coming soon for training and inference.

LightOnOCR is fully differentiable and supports:

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

📓 Finetuning notebook

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

Citation