Which OCR or document-AI model should you use?
For full-document parsing today: dots.ocr-1.5 (82.7% on our own olmOCR-bench runs) for best quality, GLM-OCR (75.7%) for value per GPU-dollar, and Tesseract only for clean printed text. Measured by us, not quoted from vendors.
24 models tracked across 19 open-source and 5 cloud services
Decision guide
Which document AI / OCR route?
Hosted API vs local OCR vs vision LLM vs hybrid, weighed by volume, privacy, validation, and evidence. Invoices as the worked example.
Original research
11 OCR models benchmarked
Our own olmOCR-Bench reproduction runs, with methodology notes and failure patterns.
Reference
Every tracked model
All OCR and document-extraction models we follow, split by open-source and cloud deployment.
Our olmOCR-bench runs
7,010 pass/fail checks across 1,403 PDFs: content preserved, reading order kept, tables and math intact. Official evaluator, our GPUs. The methodology story: why evaluation setup moves scores more than model choice.
| # | Model | Overall | Excl. baseline | Official | Notes | |
|---|---|---|---|---|---|---|
| 1 | dots.ocr-1.5 | 82.7% | — | — | Best score in our runs; layout-aware parser pipeline | |
| 2 | olmOCR-2-7B | 82.2% | 79.7% | 82.4% | The benchmark creator's own model, our closest reproduction target | |
| 3 | Qwen3.5-4B | 79.1% | — | — | General-purpose VLM, better than expected | |
| 4 | DeepSeek-OCR-2 | 78.6% | 75.6% | — | Promising quality, expensive in wall-clock time | |
| 5 | PaddleOCR-VL-1.5 | 77.3% | 74.2% | 80.0% | Respectable only once we switched to the official pipeline | |
| 6 | GLM-OCR | 75.7% | 72.2% | 75.2% | Excellent value for a 0.9B model | |
| 7 | LightOnOCR-2-1B | 75.5% | 72.0% | — | Strong small model, competitive with GLM-OCR | |
| 8 | Qwen3.5-2B | 72.5% | — | — | Better than the tiny model, still not an OCR specialist | |
| 9 | Qwen3.5-0.8B | 57.6% | — | — | Good on some tasks, clearly not enough on full-doc parsing | |
| 10 | Tesseract 5.5.2 | 34.1% | 24.8% | — | Fine for plain text, non-competitive for math/tables | |
| 11 | GOT-OCR 2.0 | 27.6% | 17.7% | — | Poor fit for this benchmark |
Key decisions
Choosing a document AI stack is a workflow decision, not a leaderboard pick. These are the tradeoffs that actually move cost and reliability; the model tables below are the evidence.
Extraction approach
OCR engine or vision LLM?
Vision LLMs handle messy layouts flexibly, but you pay per token on every page. Dedicated OCR is far cheaper at volume and often accurate enough. Test both on your worst scans before committing.
Compare document AI / OCR routes→Deployment
Hosted API or open-source on-prem?
Open-source removes per-page fees and keeps data on your network, but you own the GPU, the scaling, and the ops. It pays off above a volume threshold, or when compliance requires it.
Compare open-source vs cloud→Evidence
How accurate are these models, really?
Vendor numbers come from clean benchmarks. We reran olmOCR-Bench ourselves across 11 models; the failure patterns matter more than the headline score.
See our benchmark runs→Reliability
Where should human review sit?
A reliable cost-aware path routes only low-confidence extractions to a person. A validation gate plus a smaller model usually beats paying for a bigger one on every document.
See the worked example→Model & vendor appendix
Every tracked OCR and document-extraction model, split by deployment type. Reference material for the decisions above, not a ranking to read top to bottom.
Open source models
| Model | Provider | olmOCR | Capabilities |
|---|---|---|---|
| Tesseract | — | OCR | |
| PaddleOCR Baidu | Baidu | 80 | OCRLayout AnalysisTable Extraction |
| EasyOCR Jaided AI | Jaided AI | — | OCR |
| Surya Datalab | Datalab | — | OCRLayout Analysis |
| Marker Datalab | Datalab | 76.1 | OCRLayout AnalysisTable ExtractionDocument Conversion |
| Docling IBM Research | IBM Research | 76.9 | Layout AnalysisTable ExtractionDocument Conversion |
| GOT-OCR 2.0 StepFun | StepFun | 73.6 | OCRLayout AnalysisTable Extraction |
| MinerU OpenDataLab | OpenDataLab | 75.2 | OCRLayout AnalysisTable ExtractionDocument Conversion |
| olmOCR 2 Allen AI | Allen AI | 82.4 | OCRLayout Analysis |
| DeepSeek-OCR DeepSeek | DeepSeek | 75.7 | OCR |
| HunyuanOCR Tencent | Tencent | — | OCRTable ExtractionData Extraction |
| MiniCPM-V 4.5 OpenBMB | OpenBMB | — | OCRLayout AnalysisTable ExtractionData Extraction |
| Qwen3-VL Alibaba | Alibaba | 64.3 | OCRLayout AnalysisTable ExtractionData Extraction |
| InternVL 3.5 OpenGVLab | OpenGVLab | — | OCRLayout AnalysisTable ExtractionData Extraction |
| LightOnOCR-2-1B LightOn | LightOn | 83.2 | OCRLayout AnalysisTable ExtractionDocument Conversion |
| LayoutLM Microsoft | Microsoft | — | Layout AnalysisData ExtractionClassification |
| Donut NAVER | NAVER | — | OCRData ExtractionClassification |
| RapidOCR RapidAI | RapidAI | — | OCR |
| PP-Structure Baidu | Baidu | — | OCRLayout AnalysisTable ExtractionData Extraction |
Cloud services
| Model | Provider | olmOCR | Capabilities |
|---|---|---|---|
| Azure Document Intelligence Microsoft | Microsoft | — | OCRLayout AnalysisTable ExtractionData Extraction |
| Google Document AI | — | OCRLayout AnalysisTable ExtractionData Extraction | |
| Amazon Textract AWS | AWS | — | OCRTable ExtractionForm Processing |
| ABBYY FineReader ABBYY | ABBYY | — | OCRLayout AnalysisDocument Conversion |
| Mistral OCR 3 Mistral AI | Mistral AI | 79.1 | OCRLayout AnalysisTable ExtractionData Extraction |
Understanding Document Processing
Document processing encompasses technologies that extract structured data from unstructured documents like PDFs, images, and scanned files. The field includes several overlapping capabilities:
- OCR (Optical Character Recognition) - Extracts text from images and scanned documents
- Layout Analysis - Identifies document structure including paragraphs, headers, and columns
- Table Extraction - Reconstructs table structure with row/column relationships
- Document Conversion - Transforms documents into structured formats like Markdown or JSON
Modern approaches range from traditional OCR engines like Tesseract to AI-powered pipelines that combine vision models with language understanding. Cloud services offer managed solutions with prebuilt extractors, while open-source tools provide flexibility for custom deployments.
What changed
- 2026-07-18Page rebuilt around our own benchmark runs: full run table on the page, provenance details, and a direct answer up top.
- 2026-05-22Headline figures re-verified against the canonical run index; olmOCR-2-7B canary re-traced (0.822 vs official 0.824).
- 2026-03-21Run set extended to 15 runs across 11 models; dots.ocr-1.5 takes the lead at 82.7%.