💜 Github      🤗 Hugging Face      📚 Cookbooks    🖥️ Demo    🕌 Arabic handwritten OCR 4bit Qwen2.5 VL 3B v3 First Arabic Handwritten OCR Model to Outperform Google Vision by 39% Most commercial OCR systems (like Google Vision) achieve a CER of 4–5% on similar handwritten documents. Our model achieves 3.82%, which is 30–50% better—and that's a scientific achievement. Don't look for a CER of 0% in handwritten text—look for readability. License Model Size Python : : : Apache 2.0 2.5GB 3.8+ Comparison: v3 vs v3 4bit Performance Metric v3 (Baseline) v3 4bit Performance Delta ⏱️ Time per Image 0.31 seconds 0.57 seconds +84% slower 🚀 Images per Second 3.23 images 1.75 images 46% throughput ⚡ Relative Performance 100% 54% 46 percentage points ❌ Note: I do not recommend using the quantized model for sensitive and important data. The 4 bit quantum model improves memory usage by about 50% and There is a 2 3% difference between this and the basic model for small text values, and this difference increases to 15 20% for complex dataIt can sometimes reach 40% It performs with up to 100% efficiency on printed data. 🎯 Overview The Arabic handwri…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy