# OCRBench ![Category](https://img.shields.io/badge/Category-Reasoning-1f6feb) ![Metric](https://img.shields.io/badge/Metric-%25%20Accuracy-2da44e) ![Tasks](https://img.shields.io/badge/Tasks-1000-8957e5) ![Saturation](https://img.shields.io/badge/Saturation-Medium-d29922) ![Created](https://img.shields.io/badge/Created-May%202023-555) [![Paper](https://img.shields.io/badge/Paper-arXiv%202305.07895-b31b1b)](https://arxiv.org/abs/2305.07895) [![GitHub](https://img.shields.io/badge/GitHub-Yuliang--Liu%2FMultimodalOCR-181717?logo=github&logoColor=white)](https://github.com/Yuliang-Liu/MultimodalOCR) [![Dataset](https://img.shields.io/badge/Dataset-HuggingFace-ff9d00?logo=huggingface&logoColor=white)](https://huggingface.co/datasets/echo840/OCRBench) > **Quick answer:** OCRBench (Liu et al., 2023) is a comprehensive benchmark for OCR-related capabilities of large multimodal models, spanning text recognition, scene-text VQA, document VQA, handwritten mathematical expression recognition, and key information extraction across roughly 1,000 question-answer pairs. North Micro Vision scores 79.2% as of August 2026. ## At a Glance **What it tests:** Five OCR-centric capability groups — text recognition, scene-text-centric VQA, document-oriented VQA, key information extraction, and handwritten mathematical expression recognition. **Why it matters:** OCR was one of the first capabilities shown to be surprisingly weak in early large multimodal models despite strong general VQA performance; OCRBench was built specifically to surface and track this gap. **Known limitations:** As one of the earlier dedicated OCR benchmarks, it is now showing medium saturation among leading multimodal models, motivating the harder, larger-scale OCRBench v2 follow-up. ## What OCRBench Measures OCRBench assembles roughly 1,000 question-answer pairs across five OCR-related capability groups: plain text recognition, scene-text VQA (reading text embedded in natural photos), document-oriented VQA, key information extraction (e.g. pulling structured fields from receipts or forms), and handwritten mathematical expression recognition. The benchmark's authors found that even multimodal LLMs performing well on general VQA benchmarks often struggled significantly with precise text recognition and extraction — motivating a benchmark that isolates OCR capability from broader visual reasoning. OCRBench remains a widely-reported OCR sanity check, though its narrower five-task-group scope and smaller question count (relative to the newer, larger OCRBench v2) means it now discriminates less sharply among the strongest current models. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Reasoning (OCR) | | Metric | % accuracy (normalized; original scale 0-1000) | | Number of tasks | ~1,000 | | Saturation | Medium | | Created by | Liu et al. | | Source paper | [Liu et al. 2023](https://arxiv.org/abs/2305.07895) | | GitHub | [Yuliang-Liu/MultimodalOCR](https://github.com/Yuliang-Liu/MultimodalOCR) | | Dataset | [echo840/OCRBench on HuggingFace](https://huggingface.co/datasets/echo840/OCRBench) | ## How OCRBench Is Scored Each of the ~1,000 items across the five capability groups is scored for correctness; results are commonly reported on a 0-1000 raw point scale or normalized to a 0-100% accuracy scale, summed/averaged across all five task groups. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [North Micro Vision Instruct](/models/cohere-labs/north-micro-vision-instruct) | 79.2% | [North Micro Vision launch blog](https://huggingface.co/blog/CohereLabs/meet-north-micro-vision-instruct) | 2026-08 | *Score sourced from Cohere Labs' North Micro Vision Instruct launch announcement, August 2026.* ## OCRBench on Benchgen > No Benchgen results yet — [be the first to run OCRBench](/benchmarks/liu-et-al/ocrbench). ## OCRBench vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **OCRBench** | Five OCR capability groups: recognition, scene/doc VQA, KIE, handwriting | Medium | | [OCRBench v2](/benchmarks/fu-et-al/ocrbench-v2) | Larger-scale, harder OCR benchmark with text localization and reasoning | Low | | [DocVQA](/benchmarks/cvit-iiit-hyderabad/docvqa) | Scanned/plain-text document question answering | High | OCRBench remains a useful, widely-adopted OCR baseline, but OCRBench v2's larger and more difficult task set is the better choice for differentiating among today's strongest OCR-capable models. ## Run OCRBench on Your Model Benchgen lets you evaluate your own vision-language model's OCR capabilities across text recognition, scene/document VQA, and handwriting recognition, and track results across model versions. ## Related Models & Benchmarks - Top performers: [North Micro Vision Instruct](/models/cohere-labs/north-micro-vision-instruct) - Related benchmarks: [OCRBench v2](/benchmarks/fu-et-al/ocrbench-v2) · [DocVQA](/benchmarks/cvit-iiit-hyderabad/docvqa) ## Frequently Asked Questions

What is OCRBench?

OCRBench is a benchmark testing OCR-related capabilities of large multimodal models across text recognition, scene/document VQA, key information extraction, and handwritten math recognition.

What does a good OCRBench score look like?

North Micro Vision Instruct reports 79.2% as of August 2026; the strongest OCR-specialist models can exceed 85-90%.

Who created OCRBench?

OCRBench was created by Liu et al.; see the [original paper](https://arxiv.org/abs/2305.07895).

Is OCRBench saturated?

It shows medium saturation among leading models, which motivated the creation of the harder OCRBench v2 follow-up benchmark.

OCRBench — Results

OCRBench — Phases

OCRBench