Performance Benchmarks & Comparisons
Rigorous testing on standard layouts and long document benchmarks demonstrates that Unlimited OCR delivers high accuracy, fast processing speeds, and low memory consumption.
OmniDocBench Accuracy Metrics
The OmniDocBench dataset is a standardized testing framework designed to evaluate document parsing systems. It includes thousands of test pages containing complex columns, equations, charts, and diverse layouts.
| Model Name | Active Parameters | OmniDocBench v1.5 Score | OmniDocBench v1.6 Score |
|---|---|---|---|
| Unlimited OCR | 500 Million (3B Total) | 93.0% | 93.8% |
| DeepSeek OCR | 500 Million (3B Total) | 87.0% | 90.1% |
| Hunyuan OCR | 3 Billion | 88.2% | 89.5% |
| Got OCR | 3 Billion | 85.4% | 86.9% |
| Qwen2-VL-72B | 72 Billion (Dense) | 89.0% | 91.3% |
As shown in the table, Unlimited OCR outperforms both its predecessor baseline and other competing models. In spite of being a 3B parameter Mixture of Experts architecture (only using 500M active parameters during execution), it achieves higher scores than general purpose vision-language models like Qwen2-VL-72B on document parsing tasks.
Hardware Memory Footprint
For local execution, video memory consumption is a critical factor. Unlimited OCR runs on consumer-grade graphics processing hardware.
| Operational State | VRAM Consumption | Speed Advantage (vs. Baseline) |
|---|---|---|
| Idle Model Loaded | 6.8 GB | Constant Startup |
| Active Stream (1-5 Pages) | 7.5 GB | 13% Faster |
| Active Stream (10-50 Pages) | 8.2 GB | 45% Faster (and widening) |
| Active Stream (100+ Pages) | 8.5 GB | Crashes standard models / Unlimited Runs Stable |
Speed and Throughput Analysis
On standard, single-page documents, Unlimited OCR processes text roughly 13% faster than baseline models because of its optimized attention routing.
However, the performance gap increases significantly as document lengths expand. On a 14-page document containing roughly 41,000 characters of text, academic formulas, and tables, Unlimited OCR completes the entire extraction in a single pass in under three minutes, using less than 8.5 gigabytes of VRAM. Standard models degrade in speed or experience VRAM memory exceptions when attempting similar tasks.
This stable speed makes Unlimited OCR the standard choice for batch processing operations, library digitization, and automated document ingestion pipelines.
Data verified on standard Ubuntu environments using Nvidia RTX graphics cards. Detailed configuration specifications and inference scripts are available on the Official GitHub Repository.