OCR & Model Comparison

Independent local test results

98.3%OCR completion · 176/179

74.3%Historical end-to-end completion · 133/179

100%Latest stabilized runs · 53/53

Five full runs are included. Documents overlap between runs, so 179 is a processing-event count, not 179 unique files. Weighted model confidence for successful classifications is about 92.0%; confidence is not ground-truth accuracy.

Benchmark-informed overall score

Value axis: overall score (/10) · Category axis: tool/model · utilities excluded · sources current to 24 August 2026

Detailed comparison

10/10 = strongest fit within this comparison, not universal perfect accuracy · 8–9 = very strong · 6–7 = situational · – = not applicable

Tool / Model	OCR	Document Classification	Layout / Tables	Bounding Boxes	JSON / Structured Output	Local / Self-hosted	Speed / Efficiency	Adaptability	Production Readiness	Overall

Interpretation and corrections

The previous overall values were not consistently reproducible from the category values. Here, every overall value is calculated automatically. OpenCV, EmbeddingGemma and PaddleX Custom Classifier are rated only for the role they can actually perform.

The current local test stack scores 8.7/10 as a practical local pipeline. Its main remaining deduction is the absence of expert-confirmed labels and production-scale monitoring, not OCR completion.

Sources and evidence scope

PaddleOCR PP-StructureV3 documentation and benchmark — layout, reading order and inference performance.
PaddleOCR-VL 1.6 documentation — official 96.3% OmniDocBench v1.6 claim.
Qwen3-VL official repository — OCR, document parsing and grounding capabilities.
Gemini 3.5 Flash-Lite documentation and Gemma 4 documentation.
Azure Document Intelligence Layout documentation.
Local evidence: archived independent test reports from 18–23 August 2026. No external model was rerun on the private document set.