Updated March 2026#1 OCR ResourceIndependent & Verified

Best OCR Models 2026

Compare 158+ models across 86 benchmarks. Open source vs vendor APIs. Real results from OmniDocBench, OCRBench v2, and olmOCR — independently verified.

158+

Models Tracked

86

Benchmarks

349

Verified Results

92.86

SOTA (OmniDoc)

Quick Answers

Best for document parsing:
PaddleOCR-VL 7B

92.86 OmniDocBench — open source

Best for pure text extraction:
GPT-4o

0.02 edit distance — API

Best open-source all-rounder:
dots.ocr 3B

88.41 composite — 100+ languages

Best for Chinese documents:
Gemini 2.5 Pro

62.2% OCRBench v2 Chinese

Best lightweight / edge:
PaddleOCR-VL 0.9B

Near-top accuracy, tiny footprint

Best free OCR library:
PaddleOCR

Apache 2.0, leads all OS benchmarks

What are you trying to extract?

Pick your document type. See what actually works.

Invoices & Receipts

Line items, totals, vendor info → structured data

Best: Docling (free, local)

Handwritten Notes

Forms, signatures, meeting notes, historical docs

Best: GPT-4o (API)

PDFs & Reports

Multi-page documents, layout, tables, headers

Best: Docling (free, local)

Photos & Screenshots

Camera captures, screen grabs, social media

Best: PaddleOCR (free, local)

Scanned Books

Digitize printed text, old documents, archives

Best: PaddleOCR-VL (free, local)

ID Cards & Passports

KYC verification, identity documents, MRZ codes

Best: Azure/Google (enterprise)

Performance at a Glance

Visual comparison of accuracy, cost, and cross-benchmark coverage.

Accuracy — Top 10 Models (OmniDocBench)

Horizontal bar chart comparing top 10 OCR models by OmniDocBench composite score. PaddleOCR-VL leads at 92.86.

Cost — Price per 1,000 Pages

Cost comparison chart showing vendor APIs at $1-15 per 1K pages versus self-hosted PaddleOCR-VL at $0.09.

Cross-Benchmark Heatmap — 12 Models × 8 Benchmarks

Heatmap showing OCR model performance across 8 benchmarks. Green = high score. PaddleOCR-VL and Gemini 2.5 Pro show broadest coverage.

We Run Our Own Benchmarks

No vendor claims. Real results. Independently verified.

Full datasets. Official evaluation tools. Reproducible results. 1,355 images processed at $2.71 total cost.

dots.ocr 3BOSS
Composite
88.41
Text
95.2%
Open source →
Mistral OCR 3VERIFIED
Composite
79.75
Text
90.1%
Full results →
clearOCRVERIFIED
Composite
31.7
Text
84.6%
Traditional OCR →

Decision Tools

Model Comparator

Interactive side-by-side. Select 2–4 models, see failure modes.

Decision Guide

Failure taxonomy, decision matrix, what breaks in production.

Enterprise Toolkit

RFP templates, procurement checklists, vendor risk registers.

Vendor Partners

Get independently benchmarked. Build enterprise trust.

Open Source OCR Benchmark

Run on your own servers. No API costs. Full data privacy.

ModelOmniDocBenchOCRBench (EN)olmOCRLicense
PaddleOCR-VL
Baidu
92.86—80.0Apache 2.0
PaddleOCR-VL 0.9B
Baidu
92.56——Apache 2.0
MinerU 2.5
OpenDataLab
90.67—75.2AGPL-3.0
Qwen3-VL-235B
Alibaba
89.15——Qwen License
MonkeyOCR-pro-3B
Unknown
88.85——Apache 2.0 / MIT
OCRVerse 4B
Unknown
88.56——Apache 2.0 / MIT
dots.ocr 3B
RedNote HILab
88.41—79.1Apache 2.0
Qwen2.5-VL
Alibaba
87.02——Apache 2.0
Chandra v0.1.0
datalab-to
——83.1Apache 2.0
Infinity-Parser 7B
Unknown
——82.5Apache 2.0 / MIT
olmOCR v0.4.0
Allen AI
——82.4Apache 2.0
Marker 1.10.0
VikParuchuri
——76.5Apache 2.0 / MIT
Marker 1.10.1
VikParuchuri
——76.1Apache 2.0 / MIT
DeepSeek OCR
DeepSeek
——75.4Apache 2.0 / MIT
GPT-4o (Anchored)
OpenAI
——69.9Apache 2.0 / MIT
Nanonets OCR2 3B
Nanonets
——69.5Apache 2.0 / MIT
Gemini Flash 2
Google
——63.8Apache 2.0 / MIT
Qwen3-Omni-30B
Alibaba
—61.3%—Qwen License
Nemotron Nano V2 VL
NVIDIA
—61.2%—NVIDIA Open Model License
GPT-4o Mini
OpenAI
—44.1%—Apache 2.0 / MIT
CoCa (finetuned)
Google
———Apache 2.0
ViT-G/14
Google
———Apache 2.0
ViT-H/14
Google
———Apache 2.0
ViT-L/16
Google
———Apache 2.0
ViT-B/16
Google
———Apache 2.0
ConvNeXt V2 Huge
Meta
———MIT
ConvNeXt V2 Base
Meta
———MIT
ConvNeXt V2 Tiny
Meta
———MIT
Swin Transformer V2 Large
Microsoft
———MIT
Swin Transformer Large
Microsoft
———MIT
EfficientNetV2-L
Google
———Apache 2.0
EfficientNet-B7
Google
———Apache 2.0
EfficientNet-B0
Google
———Apache 2.0
DeiT-B Distilled
Meta
———Apache 2.0
DeiT-B
Meta
———Apache 2.0
ResNet-152
Microsoft
———MIT
ResNet-50
Microsoft
———MIT
ResNet-50 (A3 training)
Timm
———Apache 2.0
Qwen2.5-VL 72B
Alibaba
———Apache 2.0
CHURRO (3B)
Stanford
———Apache 2.0 / MIT
InternVL2-76B
Shanghai AI Lab
———MIT
InternVL3-78B
Shanghai AI Lab
———Apache 2.0 / MIT
Tesseract
Google (Open Source)
———Apache 2.0
EasyOCR
JaidedAI
———Apache 2.0
Gemini 2.5 Flash
Google
———Apache 2.0 / MIT
olmOCR v0.3.0
Allen AI
———Apache 2.0 / MIT
Qwen2-VL 72B
Alibaba
———Apache 2.0 / MIT
Qwen2.5-VL 32B
Alibaba
———Apache 2.0 / MIT
AIN 7B
Research
———Apache 2.0 / MIT
Azure OCR
Microsoft
———Apache 2.0 / MIT
PaddleOCR
Baidu
———Apache 2.0 / MIT
InternVL3 14B
OpenGVLab
———Apache 2.0 / MIT
o1-preview
OpenAI
———Apache 2.0 / MIT
Llama 3 70B
Meta
———Apache 2.0 / MIT
DeepSeek V3
DeepSeek
———Apache 2.0 / MIT
DeepSeek V2.5
DeepSeek
———Apache 2.0 / MIT
Claude 3.5 Opus
Anthropic
———Apache 2.0 / MIT
AL-Negat
Research
———Apache 2.0 / MIT
GCN
Research
———Apache 2.0 / MIT
Multi-Task Transformer
Research
———Apache 2.0 / MIT
Deep Learning (Heinsfeld)
Research
———Apache 2.0 / MIT
PHGCL-DDGFormer
Research
———Apache 2.0 / MIT
Random Forest
Baseline
———Apache 2.0 / MIT
MAACNN
Research
———Apache 2.0 / MIT
Multi-Atlas DNN
Research
———Apache 2.0 / MIT
Abraham Connectomes
Research
———Apache 2.0 / MIT
Go-Explore
Uber AI
———Apache 2.0 / MIT
BrainGNN
Research
———MIT
MVS-GCN
Research
———Apache 2.0 / MIT
BrainGT
Research
———Apache 2.0 / MIT
SVM with Connectivity Features
Research
———Apache 2.0 / MIT
AE-FCN
Research
———Apache 2.0 / MIT
DeepASD
Research
———Apache 2.0 / MIT
MCBERT
Research
———Apache 2.0 / MIT
ASD-SWNet
Research
———Apache 2.0 / MIT
Agent57
DeepMind
———Apache 2.0 / MIT
MuZero
DeepMind
———Apache 2.0 / MIT
DreamerV3
DeepMind
———Apache 2.0 / MIT
Rainbow DQN
DeepMind
———Apache 2.0 / MIT
DQN (Human-level)
DeepMind
———Apache 2.0 / MIT
Human Professional
Biology
———Apache 2.0 / MIT
BBOS-1
Unknown
———Apache 2.0 / MIT
GDI-H3
Research
———Apache 2.0 / MIT
Plymouth DL Model
Research
———Apache 2.0 / MIT
Co-DETR (Swin-L)
Research
———Apache 2.0 / MIT
InternImage-H
Shanghai AI Lab
———Apache 2.0 / MIT
DINO (Swin-L)
Research
———Apache 2.0 / MIT
YOLOv10-X
Tsinghua
———Apache 2.0 / MIT
Mask2Former (Swin-L)
Meta
———Apache 2.0 / MIT
EfficientDet-D7x
Google
———Apache 2.0 / MIT
CheXNet
Stanford ML Group
———MIT
TorchXRayVision
Cohen Lab
———Apache 2.0
CheXzero
Harvard/MIT
———MIT
MedCLIP
Research
———MIT
GLoRIA
Stanford
———MIT
BioViL
Microsoft
———MIT
RAD-DINO
Microsoft
———MIT
CheXpert AUC Maximizer
Stanford
———Apache 2.0 / MIT
DenseNet-121 (Chest X-ray)
Research
———MIT
ResNet-50 (Chest X-ray)
Research
———MIT
ConVIRT
NYU
———Apache 2.0 / MIT
PatchCore
Amazon
———Apache 2.0
PaDiM
Research
———Apache 2.0
FastFlow
Research
———MIT
EfficientAD
Research
———MIT
SimpleNet
Research
———MIT
DRAEM
Research
———MIT
CFLOW-AD
Research
———Apache 2.0
Reverse Distillation
Research
———MIT
YOLOv8 (Weld Detection)
Ultralytics
———AGPL-3.0
DefectDet (ResNet)
Research
———Apache 2.0 / MIT
DeepSeek-R1
DeepSeek
———Apache 2.0 / MIT
Llama 3.1 405B
Meta
———Apache 2.0 / MIT
Llama 3.1 70B
Meta
———Apache 2.0 / MIT
When to use open source:
  • • Sensitive data that can't leave your network
  • • High volume processing (no per-page costs)
  • • Offline / air-gapped environments
  • • Full control over the pipeline

Vendor API Benchmark

Pay per page. Fast to integrate. Enterprise support available.

VendorOmniDocBenchOCRBench (EN)olmOCRPrice/1k pages
Gemini 2.5 Pro
Google
88.0359.3%—varies
Mistral OCR 3
Mistral
79.75—78.0varies
Mistral OCR 2
Mistral
——72.0varies
Seed1.6-vision
ByteDance
—62.2%—varies
GPT-4o
OpenAI
—55.5%—varies
Claude Sonnet 4
Anthropic
—42.4%—varies
clearOCR
TeamQuest
31.70——varies
Gemini 2.0 Flash
Google
———varies
Gemini 1.5 Pro
Google
———varies
Claude 3.5 Sonnet
Anthropic
———varies
o1
OpenAI
———varies
o1-mini
OpenAI
———varies
o3
OpenAI
———varies
o3-mini
OpenAI
———varies
o4-mini
OpenAI
———varies
GPT-4.1
OpenAI
———varies
GPT-4.5 Preview
OpenAI
———varies
Claude 3.7 Sonnet
Anthropic
———varies
Grok 2
xAI
———varies
Claude 3 Opus
Anthropic
———varies
GPT-4 Turbo
OpenAI
———varies
Claude Sonnet 4.5
Anthropic
———varies
When to use vendor APIs:
  • • Need reasoning / context understanding (GPT-4o, Gemini)
  • • Low volume, occasional use
  • • Need enterprise SLA / support
  • • No infrastructure to maintain

Cross-Benchmark Champions

Models that perform well across multiple OCR benchmarks — not just one.

Gemini 2.5 Pro#1
5 benchmarks3 top-3avg #3.4
OmniDoc: #8OCRBench: #4CHURRO: #2VideoOCR: #1Thai: #2
PaddleOCR-VL#2
2 benchmarks1 top-3avg #2.5
OmniDoc: #1olmOCR: #4
Gemini 1.5 Pro#3
2 benchmarks1 top-3avg #3.0
CC-OCR: #1VideoOCR: #5
Qwen2.5-VL 72B
2 benchmarks1 top-3avg #3.5
VideoOCR: #2Thai: #5
GPT-4o
4 benchmarks1 top-3avg #4.3
OCRBench: #6CC-OCR: #4VideoOCR: #4Arabic: #3
Qwen2.5-VL 32B
2 benchmarks1 top-3avg #4.5
VideoOCR: #6Thai: #3
MinerU 2.5
2 benchmarks1 top-3avg #7.0
OmniDoc: #3olmOCR: #11
Claude Sonnet 4
2 benchmarks1 top-3avg #9.0
OCRBench: #17Thai: #1

Deep Dives & Techniques

The OCR Economics Shift

167:1

Self-hosted VLM-OCR is now better AND 167× cheaper than vendor APIs. The October 2025 inflection point.

How Docling Works

Architecture of IBM's document understanding library. Why VLM pipelines outperform traditional OCR.

Interactive OCR Correction

Handling OCR “flicker” (H vs N) and camera drift in mobile apps. Google MLKit + centroid anchoring.

OCR Benchmarks Directory

26 benchmarks across document parsing, handwriting, video OCR, scene text, and multilingual tasks.

Rys OCR

POLISH SOTA

71% CER reduction on Polish diacritics. LoRA fine-tune of PaddleOCR-VL. Apache 2.0.

Implementation Guide

You Know the Best OCR Model. Now Ship It.

3 questions, one recommendation, copy-paste code that runs in 10 minutes.

Get started
#1 on OmniDocBench

Run this OCR on your Mac — $25, one-time

Hardparse runs PaddleOCR-VL locally via Metal. No cloud, no subscription. Tables, formulas, 109 languages.

Get HardparseLearn more

All OCR Content

Model Reviews

Comparisons

Guides & Tools

Frequently Asked Questions

What is the best OCR model in 2026?
PaddleOCR-VL 7B leads the OmniDocBench leaderboard with a composite score of 92.86, outperforming GPT-4o and Gemini 2.5 Pro on end-to-end document parsing including tables and formulas. It's open source (Apache 2.0).
Which OCR model has the best English text recognition?
On OCRBench v2, Seed1.6-vision leads English OCR at 62.2%, followed by Qwen3-Omni-30B at 61.3%. For pure text extraction, GPT-4o achieves the lowest edit distance (0.02) on OmniDocBench.
Is open-source OCR better than paid APIs in 2026?
Yes. PaddleOCR-VL (open source) scores 92.86 on OmniDocBench vs GPT-4o's 85.80. Self-hosted VLM-OCR is 167× cheaper per page than vendor APIs while delivering higher accuracy on document parsing tasks.
Which OCR is best for invoices and receipts?
PaddleOCR-VL excels at structured documents with its 93.52 TEDS score on table recognition. For managed solutions, Docling (IBM, free) or Google Document AI offer strong invoice-specific pipelines.
How much does OCR cost per page?
Vendor APIs range from $1–$15 per 1,000 pages (Mistral OCR at $1, GPT-4o at ~$15). Self-hosted PaddleOCR-VL costs approximately $0.09 per 1,000 pages on a consumer GPU — a 167× cost reduction.

Have benchmark results?

Submit your paper or results. We verify and add them to our database.

Submit Paper

Get OCR updates

New models, benchmark results, and practical guides.

About This Data

All benchmark results sourced from AlphaXiv leaderboards, published papers, and our own independent verification. Each data point includes source URL and access date.

Results marked “pending verification” are claimed in papers but not independently confirmed. We do not include estimated or interpolated values.

JSON APIMethodologyGitHub