LightOnOCR-3-4B — W4A16 Quantized (AutoRound)

This repository provides an INT4 quantized version of lightonai/LightOnOCR-3-4B using Intel AutoRound with high-fidelity production settings.

LightOnOCR-3-4B is an end-to-end vision-language model based on the Qwen3.5 architecture tailored for high-accuracy document transcription, layout understanding, visual grounding, and structured figure extraction.

---

Quantization Details

Sibling Repositories & Formats

Choose the backend format suited to your inference engine:

• AutoRound Native / Transformers: Vishva007/LightOnOCR-3-4B-W4A16-AutoRound
• AutoGPTQ / Marlin: Vishva007/LightOnOCR-3-4B-W4A16-AutoRound-GPTQ
• vLLM / Compressed Tensors: Vishva007/LightOnOCR-3-4B-W4A16-AutoRound-LLM-Compressor

---

Operating Modes

LightOnOCR-3 operates in two primary modes:

1. Plain Transcription Mode (prompt = ""): Transcribes text directly to Markdown layout without bounding boxes.
2. Grounding Mode (prompt = "grounding"): Outputs normalized bounding boxes [0, 1000] and semantic labels (text, title, table, image, chart, etc.) for each document component.

---

Quickstart

1. Python Inference via Transformers

pip install --upgrade "transformers>=5.5.4" auto-round pillow torchvision

import torch
from transformers import AutoProcessor, AutoModelForImageTextToText

# Swap repo_id with your target variant:
model_id = "Vishva007/LightOnOCR-3-4B-W4A16-AutoRound"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Set image URL or local path
image_url = "[https://huggingface.co/datasets/hf-internal-testing/fixtures_ocr/resolve/main/SROIE-receipt.jpeg](https://huggingface.co/datasets/hf-internal-testing/fixtures_ocr/resolve/main/SROIE-receipt.jpeg)"

# Use "grounding" or "" (for transcription only)
prompt = "grounding"

content = [{"type": "image", "url": image_url}]
if prompt:
    content.append({"type": "text", "text": prompt})

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

inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    enable_thinking=False,  # LightOnOCR is optimized with thinking disabled
).to(model.device)

with torch.inference_mode():
    output_ids = model.generate(**inputs, max_new_tokens=2048, do_sample=False)

generated_ids = output_ids[0, inputs["input_ids"].shape[1]:]
print(processor.decode(generated_ids, skip_special_tokens=True))

---

2. High-Throughput Serving with vLLM

(Recommended using Vishva007/LightOnOCR-3-4B-W4A16-AutoRound-LLM-Compressor or Vishva007/LightOnOCR-3-4B-W4A16-AutoRound-GPTQ)

Start Server

vllm serve Vishva007/Vishva007/LightOnOCR-3-4B-W4A16-AutoRound \
  --trust-remote-code \
  --dtype bfloat16 \
  --max-model-len 8192 \
  --default-chat-template-kwargs '{"enable_thinking": false}'

Query Server

import requests

url = "http://localhost:8000/v1/chat/completions"
image_url = "[https://huggingface.co/datasets/hf-internal-testing/fixtures_ocr/resolve/main/SROIE-receipt.jpeg](https://huggingface.co/datasets/hf-internal-testing/fixtures_ocr/resolve/main/SROIE-receipt.jpeg)"

payload = {
    "model": "Vishva007/LightOnOCR-3-4B-W4A16-AutoRound",
    "messages": [
        {
            "role": "user",
            "content": [
                {"type": "image_url", "image_url": {"url": image_url}},
                {"type": "text", "text": "grounding"}
            ]
        }
    ],
    "max_tokens": 2048,
    "temperature": 0.0
}

response = requests.post(url, json=payload)
print(response.json()["choices"][0]["message"]["content"])

---

Best Practices for Document Preprocessing

• Resolution: Render scanned documents and PDFs at 400 DPI using a target longest dimension of 2048px.
• Aspect Ratio: Maintain the native aspect ratio when resizing; do not stretch images, as bounding box coordinate accuracy depends on geometric integrity.
• Thinking Tokens: Ensure enable_thinking=False is preserved in chat templates during generation.

---

🚀 Deploy on RunPod

One-click launch environments pre-configured with PyTorch, CUDA, and dependencies for fine-tuning or quantization.

🎁 Need GPU compute? Sign up via RunPod and get $5–$500 in free credits when you add your first $10.

PyTorch 2.15

PyTorch 2.14

PyTorch 2.13

PyTorch 2.12

---

Citation & Acknowledgements

Quantized using Intel AutoRound. Base model created and released by LightOn AI.

@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/abs/2601.14251](https://arxiv.org/abs/2601.14251)}}
}

@article{cheng2023optimize,
  title={Optimize Weight Rounding via Signed Gradient Descent for Large Language Model Quantization},
  author={Cheng, Wenhua and Hai, Wei and Shen, Haihao and others},
  journal={arXiv preprint arXiv:2309.05516},
  year={2023}
}