LightOnOCR-2-1B Demo - Local Version

A local version of LightOnOCR-2-1B-Demo

Created by Claude Code | Orchestrated by TheAwakenOne

Overview

LightOnOCR-2-1B is a 1B-parameter end-to-end multilingual vision-language model that converts document images (PDFs) into clean, naturally ordered text without brittle OCR pipelines.

Based on the paper: LightOnOCR: A 1B End-to-End Multilingual Vision-Language Model for State-of-the-Art OCR

Key Capabilities

• State-of-the-art OCR: 83.2% on OlmOCR-Bench
• 9x smaller and faster than prior best-performing models
• Complex layout handling: Tables, forms, receipts, scientific notation
• Bounding box detection: Localize embedded images (bbox variants)
• LaTeX math transcription: KaTeX compatible
• Scanned document support: Robust to noise and degradation
• Multilingual: European languages and Latin scripts
• Max resolution: 1540px (longest edge)
• Throughput: 5.71 pages/sec on H100

Model Variants

How to Use

Via Pinokio (Recommended)

1. Install Pinokio
2. Search for "LightOnOCR-2-1B-Pinokio" or download this repository
3. Click "Install" to set up dependencies
4. Click "Start" to launch the app
5. Open the Web UI and upload an image or PDF

System Requirements

• GPU: NVIDIA GPU with 4GB+ VRAM (recommended) or Apple Silicon Mac
• RAM: 8GB+ system memory
• Storage: ~5GB for model weights
• Python: 3.10+

Performance Benchmarks

OlmOCR-Bench Results

Inference Throughput (H100)

Links

• Paper (arXiv)
• Blog Post
• Model Collection
• Training Dataset
• Finetuning Notebook
• Original HuggingFace Space

License

Model weights: Apache 2.0

---

A local version of lightonai/LightOnOCR-2-1B-Demo

Created by Claude Code | Orchestrated by TheAwakenOne