Best Open-Source OCR Models 2026: 5 Ranked by Benchmark
The short answer
PaddleOCR-VL-1.6 is the best all-round open OCR model as of October 2026, GLM-OCR the easiest to run, and Chandra OCR 2 the most accurate on hard scans — if its licence fits you. The 2026 generation are small vision-language models (0.9B–7B parameters) that output Markdown, HTML or JSON with tables and reading order, not the line-by-line text of older engines like Tesseract.
Two benchmarks matter and they are not comparable with each other: OmniDocBench (document parsing: text, tables, formulas, reading order) and Ai2’s olmOCR-Bench (7,000+ unit tests over 1,400 documents, including old scans and tiny text). All scores below are reported by the model’s own authors.
The ranking
| # | Model | Maker / release | Size | Licence | Headline score | Best for |
|---|---|---|---|---|---|---|
| 1 | PaddleOCR-VL-1.6 | Baidu, May 28, 2026 | 0.9B | Apache 2.0 | 96.33% OmniDocBench v1.6 | Best accuracy-per-GPU, tables, formulas, seals |
| 2 | GLM-OCR | Zhipu (Z.ai), early 2026 | 0.9B | MIT | 94.62 OmniDocBench v1.5 | Easy deployment (vLLM, SGLang, Ollama), high concurrency |
| 3 | Chandra OCR 2 | Datalab, March 2026 | ~5B | Apache 2.0 code, OpenRAIL-M weights (revenue cap) | 85.8 olmOCR-Bench | Handwriting, forms, 90+ languages |
| 4 | DeepSeek-OCR 2 | DeepSeek, Jan 27, 2026 | ~3B | Apache 2.0 | — (see note) | Throughput, document-to-Markdown at scale |
| 5 | olmOCR 2 | Ai2, Oct 21, 2025 | 7B | Apache 2.0 | 82.4 olmOCR-Bench | Turnkey PDF-to-text pipeline, LLM datasets |
Why each one is ranked where it is
1. PaddleOCR-VL-1.6. Baidu’s upgrade of PaddleOCR-VL-1.5 keeps the 0.9B size and adds region-aware data refinement and RL post-training. It claims state of the art on OmniDocBench v1.6 (96.33%), v1.5 and Real5-OmniDocBench, with gains on tables, rare characters, seals and charts. It is a drop-in replacement for 1.5, ships GGUF builds, and serves through a vLLM-backed paddleocr genai_server.
2. GLM-OCR. A CogViT vision encoder plus a GLM-0.5B decoder, with a two-stage pipeline (PP-DocLayoutV3 layout, then parallel recognition). It reported 94.62 on OmniDocBench v1.5 at release. The practical advantage is tooling: an SDK, a hosted API and support for vLLM, SGLang and Ollama, so it fits existing LLM serving stacks.
3. Chandra OCR 2. Datalab’s model tops the olmOCR-Bench table it publishes (85.8 ± 0.8, ahead of dots.ocr 1.5 at 83.9, Chandra 1 at 83.1 and olmOCR 2 at 82.4), with strong handwriting, forms with checkboxes, math and a 90-language multilingual benchmark. It would rank first on hard scans; it ranks third because commercial self-hosting above $2 million in revenue or funding requires a paid licence, and the weights may not be used to compete with Datalab’s API.
4. DeepSeek-OCR 2. A ~3B “Visual Causal Flow” model that reorders visual tokens by document structure, aimed at compressing pages into few tokens for fast throughput. Apache 2.0, vLLM and Transformers inference. DeepSeek’s repository does not headline a single comparable score, so benchmark it on your own pages.
5. olmOCR 2. Ai2’s 7B model and full toolkit convert PDFs and images to clean Markdown with natural reading order, strip headers and footers, and handle tables and equations. Ai2 puts the cost at under $200 per million pages on your own GPU. Older and larger than the leaders, but the most complete pipeline.
How to choose in 30 seconds
- Commercial product, one GPU, mixed business documents: PaddleOCR-VL-1.6.
- You already run vLLM or Ollama and want the least integration work: GLM-OCR.
- Handwritten forms or many languages, and you are under $2M revenue (or will pay Datalab): Chandra OCR 2.
- Millions of pages into a RAG or training corpus: DeepSeek-OCR 2 or olmOCR 2.
- No GPUs at all: use a cloud API instead — see best OCR and document extraction APIs.
Before you trust a benchmark
- Scores are self-reported and each team picks its benchmark version; do not compare an OmniDocBench number with an olmOCR-Bench number.
- Build a 200-page test set from your own documents, with your worst scans, tables and handwriting.
- Measure what you use downstream: table structure for extraction, reading order for RAG chunking, character accuracy for search.
- Check throughput on your GPU at your batch size, not the authors’ hardware.
Related: best OCR for handwriting and complex tables, best PDF parsing API for RAG pipelines, parsing scanned PDFs with low latency.
Last verified: October 17, 2026. Licences checked on each model’s Hugging Face card.