Optical character recognition used to mean Tesseract and a lot of clean-up code. In 2026, the best OCR systems are small vision-language models that read a page the way a person does — understanding layout, tables, formulas, handwriting and reading order — and output clean Markdown, HTML or JSON. Many of the best are open-source and small enough to run on a single GPU. Here is how to choose one.
Why AI OCR beats classic OCR
Classic OCR engines recognise characters line by line. They struggle with multi-column layouts, tables, stamps, low-quality scans and mixed languages, and they do not understand structure. VLM-based OCR models:
- Preserve layout and reading order across columns and pages.
- Rebuild tables as structured data rather than jumbled text.
- Handle formulas, charts, handwriting and many languages.
- Output Markdown or JSON ready for search, RAG or databases.
That makes them the foundation for invoice processing, contract analysis, medical records, KYC documents and feeding company documents into RAG assistants.
How OCR models are benchmarked
Two benchmarks dominate:
- OmniDocBench — a broad document-parsing benchmark covering text, tables, formulas and layout. Version 1.6 was released in April 2026 (OmniDocBench).
- olmOCR-Bench — Ai2's benchmark built from unit tests on real PDFs.
On the OmniDocBench v1.6 end-to-end leaderboard (checked October 2026), the top scores were TeleOCR (96.91), OvisOCR2 (96.47), PaddleOCR-VL-1.6 (96.34), MinerU2.5-Pro (95.75) and GLM-OCR (95.22). Scores at the top are now very close; your own documents are the real test.
The leading open OCR models
PaddleOCR-VL (Baidu)
PaddleOCR-VL is a compact 0.9B-parameter model that scored 96.3% on OmniDocBench v1.6 with version 1.6, released in May 2026, and is now the default PaddleOCR pipeline (PaddleOCR docs). It supports more than 100 languages and runs comfortably on modest GPUs.
Licence: Apache 2.0. Best for: high-volume, multilingual document parsing in production.
GLM-OCR (Zhipu)
Another 0.9B model, GLM-OCR scored 95.22 on OmniDocBench v1.6 and is MIT-licensed — a strong, permissive alternative to PaddleOCR-VL.
DeepSeek-OCR 2
DeepSeek-OCR 2, released in January 2026, is a 3B-parameter model built on a new visual encoder designed to compress documents efficiently (model card). It is Apache 2.0 and particularly interesting for very long documents and for feeding pages into LLMs efficiently.
olmOCR 2 (Ai2)
olmOCR 2 is a 7B model fine-tuned from Qwen2.5-VL with reinforcement learning using unit-test rewards, scoring 82.4 on olmOCR-Bench (Ai2). It is Apache 2.0 and comes with a full open pipeline for converting large PDF collections.
dots.mocr (rednote)
Formerly dots.ocr, this roughly 3B model reports strong olmOCR-Bench results (GitHub). Note that the repository includes an MIT licence and a separate licence agreement — review both.
IBM Granite-Docling and Docling
Granite-Docling-258M is a tiny Apache 2.0 model that converts documents to IBM's structured DocTags format, used with the MIT-licensed Docling library — a popular, well-engineered toolkit for turning PDFs, Word and PowerPoint files into clean data for AI pipelines.
Chandra OCR 2 (Datalab) — check the licence
Chandra OCR 2 posted the highest olmOCR-Bench score we found among open models (85.8) and covers 90 languages (GitHub). But its weights are not freely usable by most businesses: they are free for research, personal use and startups under $2 million in funding or revenue, and cannot be used to compete with Datalab's API. Commercial self-hosting needs a paid licence. Datalab's Marker and Surya tools carry similar restrictions.
General vision-language models
Large VLMs such as Qwen3-VL (OCR in 32 languages) and Gemma 4 also read documents well and can answer questions about them directly — useful when you need understanding, not just extraction. See open-source vision-language models.
Tesseract
Tesseract 5.5 (Apache 2.0) is still maintained and useful for simple printed text on CPU-only machines, but it is no longer competitive on complex documents.
Comparison
| Model | Size | Licence | Strength |
|---|---|---|---|
| PaddleOCR-VL 1.6 | 0.9B | Apache 2.0 | Top-tier accuracy, multilingual, efficient |
| GLM-OCR | 0.9B | MIT | Top-tier accuracy, permissive |
| DeepSeek-OCR 2 | 3B | Apache 2.0 | Efficient long documents |
| olmOCR 2 | 7B | Apache 2.0 | Large PDF collections, open pipeline |
| dots.mocr | ~3B | MIT + separate agreement | Strong benchmarks |
| Granite-Docling + Docling | 258M | Apache 2.0 / MIT | Office files to structured data |
| Chandra OCR 2 | — | Restricted weights | Highest olmOCR-Bench score |
| Tesseract 5.5 | — | Apache 2.0 | CPU-only, simple text |
How to choose
- Test on your own documents. Collect 50–100 representative pages — your invoices, forms and scans — and compare outputs. Benchmark leaders differ by only a point or two.
- Check the licence for commercial use. PaddleOCR-VL, GLM-OCR, DeepSeek-OCR 2, olmOCR 2 and Granite-Docling are permissive; some top performers are not.
- Decide on output format. Markdown for RAG and search; JSON with defined fields for invoices and forms.
- Plan for validation. Even the best OCR makes mistakes. Validate totals, dates and IDs with rules, and route low-confidence documents to a human.
- Consider data privacy. Self-hosted open models keep sensitive documents — medical, financial, identity — inside your infrastructure.
A typical document-AI pipeline
- Ingest files from email, uploads or storage.
- Parse with an OCR model into Markdown or structured layout.
- Extract the fields you need with an LLM and a strict schema.
- Validate with business rules and confidence thresholds.
- Review exceptions in a simple human-in-the-loop interface.
- Send clean data into your ERP, CRM or database — or index it for an AI assistant.
How Zinetics can help
We build document-AI pipelines for invoices, contracts, medical and logistics paperwork — self-hosted for privacy or API-based for speed — integrated with the systems you already use. See how we add AI to existing apps, explore our AI development services, or send us a sample document to discuss.