PaddleOCR-VL-1.5: How a 0.9B Model Achieves 94.5% Document Parsing Accuracy

PaddleOCR-VL-1.5: The 0.9B Parameter Revolution in Document Parsing Core Question: How can a sub-1GB lightweight model achieve 94.5% accuracy in document parsing under real-world complex scenarios? The answer is straightforward: PaddleOCR-VL-1.5 delivers. This vision-language model with only 0.9B parameters achieves 94.5% accuracy on OmniDocBench v1.5, surpassing all previous comparable models. More importantly, this isn’t laboratory performance under ideal conditions—it’s real-world capability across scanning artifacts, skew, warping, screen photography, and illumination variations. My biggest takeaway from testing this model: finally, a model that understands real-world chaos. How many documents we process daily are perfectly scanned and perfectly aligned? Most are phone-captured …