image to text benchmark
OCRBench Leaderboard
OCRBench evaluates Optical Character Recognition (OCR) capabilities in Large Multimodal Models across text recognition, scene text VQA, and document understanding tasks.
Updated Sep 5, 2026
Models24
Model coverage24
MetricScore
EvidenceB
OCRBench Ranking
Higher score ranks better on this benchmark.
Sort by
Rank | Model | Score | Percentile | Participants | Evidence | Evaluated |
|---|
| Rank01 | ModelMA | Score92.30% | Percentile100.00% | Participants24 | EvidenceC | Evaluated |
| Rank02 | ModelAC | Score92.10% | Percentile95.65% | Participants24 | EvidenceC | Evaluated |
| Rank03 | ModelAC | Score92.00% | Percentile91.30% | Participants24 | EvidenceC | Evaluated |
| Rank04 | ModelAC | Score91.00% | Percentile86.96% | Participants24 | EvidenceC | Evaluated |
| Rank05 | ModelAC | Score90.30% | Percentile82.61% | Participants24 | EvidenceC | Evaluated |
| Rank06 | ModelAC | Score89.60% | Percentile78.26% | Participants24 | EvidenceC | Evaluated |
| Rank07 | ModelAC | Score89.50% | Percentile73.91% | Participants24 | EvidenceC | Evaluated |
| Rank08 | ModelAC | Score89.40% | Percentile69.57% | Participants24 | EvidenceC | Evaluated |
| Rank09 | ModelAC | Score89.40% | Percentile65.22% | Participants24 | EvidenceC | Evaluated |
| Rank10 | ModelAC | Score88.50% | Percentile60.87% | Participants24 | EvidenceC | Evaluated |
| Rank11 | ModelAC | Score88.10% | Percentile56.52% | Participants24 | EvidenceC | Evaluated |
| Rank12 | ModelAC | Score87.70% | Percentile52.17% | Participants24 | EvidenceC | Evaluated |
| Rank13 | ModelAC | Score87.50% | Percentile47.83% | Participants24 | EvidenceC | Evaluated |
| Rank14 | ModelAC | Score86.40% | Percentile43.48% | Participants24 | EvidenceC | Evaluated |
| Rank15 | ModelAC | Score85.50% | Percentile39.13% | Participants24 | EvidenceC | Evaluated |
| Rank16 | ModelMI | Score84.40% | Percentile34.78% | Participants24 | EvidenceC | Evaluated |
| Rank17 | ModelLA | Score84.20% | Percentile30.43% | Participants24 | EvidenceC | Evaluated |
| Rank18 | ModelAC | Score83.90% | Percentile26.09% | Participants24 | EvidenceC | Evaluated |
| Rank19 | ModelDE | Score83.40% | Percentile21.74% | Participants24 | EvidenceC | Evaluated |
| Rank20 | ModelAC | Score81.90% | Percentile17.39% | Participants24 | EvidenceC | Evaluated |
| Rank21 | ModelDE | Score81.10% | Percentile13.04% | Participants24 | EvidenceC | Evaluated |
| Rank22 | ModelDE | Score80.90% | Percentile8.70% | Participants24 | EvidenceC | Evaluated |
| Rank23 | ModelAC | Score80.80% | Percentile4.35% | Participants24 | EvidenceC | Evaluated |
| Rank24 | ModelCO | Score79.20% | Percentile0.00% | Participants24 | EvidenceC | Evaluated |
OCRBench Highlights
The leading models and scores on this benchmark.
OCRBench Score Distribution
A closer view of the leading scores on this benchmark.
The Top AI Models for OCRBench
The first five results on this benchmark, with official price and output speed added where the model identity can be matched.
What is OCRBench?
What OCRBench measures and how its scores work.
OCRBench is a benchmark for evaluating Optical Character Recognition (OCR) capabilities in Large Multimodal Models.
It measures performance on text recognition, scene text VQA, and document understanding tasks using the Score metric in ratio units.
Scores are shown in ratio. This benchmark is not independently verified and has an evidence level of B.
- Family
- OCRBench
- Modality
- multimodal
- Primary category
- image to text
- Score direction
- higher
- LLMBoard eligible
- No
- Evaluation key
- overall
LLM Stats. Benchmark scores retain their original unit. Overall score eligibility is shown separately.
FAQ
Common questions about OCRBench.
Which model scores highest on OCRBench?
Kimi K2.5 is currently ranked first with 92.30%.
What are the top three models on OCRBench?
The current leaders are Kimi K2.5 (92.30%), Qwen3.5-122B-A10B (92.10%), and Qwen3 VL 235B A22B Instruct (92.00%).
Which OCRBench model has the lowest official input price?
Qwen3.5-35B-A3B has the lowest matched official input price at $0.25 input / $2.0 output per 1M tokens.
Which models are fastest among OCRBench results?
The fastest matched records are Phi-4-multimodal-instruct (25.00 tok/s via DeepInfra) and Qwen3 VL 4B Thinking (8.77 tok/s via DeepInfra).
Does the highest score result prove overall model quality?
No. This benchmark measures one defined capability or task. The overall LLMBoard score uses a separate aggregation across eligible benchmark evidence.