# OCRBench v2 ![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-10%2C000-8957e5) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) ![Created](https://img.shields.io/badge/Created-Dec%202024-555) [![Paper](https://img.shields.io/badge/Paper-arXiv%202501.00321-b31b1b)](https://arxiv.org/abs/2501.00321) [![GitHub](https://img.shields.io/badge/GitHub-Yuliang--Liu%2FMultimodalOCR-181717?logo=github&logoColor=white)](https://github.com/Yuliang-Liu/MultimodalOCR) [![Leaderboard](https://img.shields.io/badge/Project-99franklin.github.io-0969da)](https://99franklin.github.io/ocrbench_v2/) > **Quick answer:** OCRBench v2 (Fu et al., 2025) is a large-scale, bilingual (English/Chinese) text-centric benchmark with roughly 4x more tasks than the original OCRBench, covering 31 diverse scenarios and 10,000 human-verified question-answer pairs — including text localization, handwritten content extraction, and logical reasoning over text-rich images. North Micro Vision scores 36.7% on the English subset as of August 2026. ## At a Glance **What it tests:** Bilingual OCR capability across 31 scenarios and 10,000 human-verified questions, emphasizing harder sub-tasks than plain text recognition — text localization, handwriting extraction, fine-grained/layout perception, complex element parsing, and logical reasoning. **Why it matters:** OCRBench v2's authors found most large multimodal models score below 50/100 on its harder task mix, showing that OCR capability claimed via simpler recognition benchmarks doesn't necessarily transfer to localization- and reasoning-heavy real-world OCR tasks. **Known limitations:** As a newer, deliberately harder benchmark, current absolute scores are still low across the board — most models, including several frontier systems, score below 50%, so results should be read as early-stage progress rather than near-ceiling performance. ## What OCRBench v2 Measures OCRBench v2 substantially expands on the original OCRBench, offering roughly 4x more task types across 31 diverse real-world scenarios, with 10,000 human-verified question-answer pairs plus a held-out private test set of 1,500 manually annotated images used to validate result consistency. Task types go well beyond plain text recognition to include precise text localization (bounding a specific piece of text within an image), handwritten content extraction, fine-grained and layout perception, complex multi-element parsing, and logical reasoning over text-rich content. The paper's central finding is that most large multimodal models — including many considered strong OCR performers on the original OCRBench — score below 50 out of 100 on this harder task mix, and share five common limitation categories: rare/uncommon text recognition, fine-grained perception, layout perception, complex element parsing, and logical reasoning. Because of this difficulty gap, OCRBench v2 is currently the more discriminative and future-proof OCR benchmark of the two, while the original OCRBench remains useful mainly as a simpler, faster baseline check. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Reasoning (OCR, text localization, layout) | | Metric | % accuracy (normalized; original scale 0-100) | | Number of tasks | 10,000 (31 scenarios) | | Saturation | Low | | Created by | Fu et al. | | Source paper | [Fu et al. 2025](https://arxiv.org/abs/2501.00321) | | GitHub | [Yuliang-Liu/MultimodalOCR](https://github.com/Yuliang-Liu/MultimodalOCR) | | Project page | [99franklin.github.io/ocrbench_v2](https://99franklin.github.io/ocrbench_v2/) | ## How OCRBench v2 Is Scored Each of the 10,000 questions is scored for correctness across the benchmark's task categories (localization, extraction, layout perception, complex parsing, and reasoning); the overall score is the percent accuracy across the full public English or Chinese test subset, with a private held-out set used to validate result reliability. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [North Micro Vision Instruct](/models/cohere-labs/north-micro-vision-instruct) | 36.7% | [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, evaluated on the English (`v2_en`) subset.* ## OCRBench v2 on Benchgen > No Benchgen results yet — [be the first to run OCRBench v2](/benchmarks/fu-et-al/ocrbench-v2). ## OCRBench v2 vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **OCRBench v2** | Bilingual, 31-scenario OCR with localization and reasoning | Low | | [OCRBench](/benchmarks/liu-et-al/ocrbench) | Five-group OCR benchmark: recognition, scene/doc VQA, KIE, handwriting | Medium | | [InfoVQA](/benchmarks/mathew-et-al/infovqa) | Dense infographics combining text, charts, and layout | Medium | OCRBench v2 is the harder, more comprehensive successor to OCRBench — use it when comparing frontier or near-frontier models where the original OCRBench no longer discriminates well. ## Run OCRBench v2 on Your Model Benchgen lets you evaluate your own vision-language model's OCR capability across text localization, handwriting extraction, and logical reasoning, 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](/benchmarks/liu-et-al/ocrbench) · [InfoVQA](/benchmarks/mathew-et-al/infovqa) ## Frequently Asked Questions

What is OCRBench v2?

OCRBench v2 is a large-scale, bilingual OCR benchmark spanning 31 scenarios and 10,000 questions, testing text localization, handwriting extraction, and logical reasoning — not just plain text recognition.

What does a good OCRBench v2 score look like?

North Micro Vision Instruct reports 36.7% on the English subset as of August 2026; most large multimodal models, including several frontier systems, score below 50%.

Who created OCRBench v2?

OCRBench v2 was created by Fu et al.; see the [original paper](https://arxiv.org/abs/2501.00321).

Is OCRBench v2 saturated?

No — it's a low-saturation, deliberately hard benchmark; most models today score well below 50 out of 100.

OCRBench v2 — Results

OCRBench v2 — Evaluations