How to choose an open-source OCR model for PDFs and scans
Compare open OCR and document-parsing models by size, license, language and tooling, and find out which one fits invoices, long PDFs, Chinese pages or a laptop.
Short answer. For most documents, start with a small parser of about 1B parameters: PaddleOCR-VL or GLM-OCR for general pages, MiniCPM-OCR for Chinese and English on a laptop or phone, LightOnOCR-3 if you also want bounding boxes, Nemotron 3.5 Parse for high-volume parsing on NVIDIA GPUs, and NuExtract3 when you want fields as JSON from a template you define. Check the license before the benchmark, because three of the twelve current OCR models on this site restrict commercial use, and test two or three candidates on your own scans, because every published score is self-reported.
What an open OCR model does today
Classic OCR finds characters in an image and returns text. The open models in this guide are small vision-language models that read a whole page and write it back as Markdown, HTML or JSON, keeping headings, tables and formulas in place. Some also return where each block sits on the page, pull fields out of forms, or translate text inside images.
Most of them have between 0.6B and 5.3B parameters, so one consumer GPU is enough. They are mostly used to turn PDFs into clean text for retrieval-augmented generation. The OCR and document models list shows every model we track, and the glossary entry OCR and document parsing explains the terms.
The shortlist
| Model | Size | License | Released | What stands out |
|---|---|---|---|---|
| PaddleOCR-VL | 0.96B | Apache 2.0 | May 2026 | Tables, formulas, charts and seals; good for Chinese-heavy pages |
| GLM-OCR | 0.9B | MIT | Feb 2026 | An SDK, plus field extraction into a fixed JSON structure |
| Xiaomi-OCR-0 | 0.9B | Apache 2.0 | Sep 2026 | A tiny first release for parsing, extraction and OCR questions |
| MiniCPM-OCR | 1.3B | Apache 2.0 | Oct 2026 | Chinese and English, photographed pages, iOS and llama.cpp support |
| HunyuanOCR | 1.1B | Tencent Hunyuan Community | Jul 2026 | Parsing, extraction and translating text inside images |
| LightOnOCR-3 | 0.8B, 1B, 4B | Apache 2.0 | Oct 2026 | Markdown with bounding boxes, figure captions and chart data |
| Nemotron 3.5 Parse | 0.63B | OpenMDW-1.1 | Oct 2026 | Text, tables, chart data, reading order and boxes; one to three passes trade accuracy for speed |
| NuExtract3 | 4.5B | Apache 2.0 | May 2026 | Fills a JSON template you define or writes Markdown, with optional reasoning |
| Unlimited-OCR | 3.3B | MIT | Jun 2026 | Dozens to 100+ pages in one pass |
| DeepSeek-OCR 2 | 3.4B | Apache 2.0 | Jan 2026 | An encoder that orders visual tokens by meaning |
| Chandra OCR 2 | 5.3B | Datalab OpenRAIL-M | Mar 2026 | 90+ languages, handwriting, forms and complex tables |
| Surya OCR 2 | 0.69B | Datalab OpenRAIL-M, US$5M limit | May 2026 | Layout, reading order and tables with a 91-language breakdown; vLLM or llama.cpp |
The first DeepSeek-OCR, from October 2025, is still downloaded a lot, but we mark it as legacy because DeepSeek-OCR 2 replaced it with a different encoder and a different license.
Pick by the kind of document
- General business documents and PDFs. Start with PaddleOCR-VL or GLM-OCR. They are about the same size, so test both on your pages; our GLM-OCR vs PaddleOCR-VL page lists what separates them.
- Invoices, IDs and forms where you need fields as JSON. GLM-OCR extracts fields into a fixed structure, MiniCPM-OCR offers key information extraction for invoices, forms and seals, and NuExtract3 fills a JSON template that you define, with null for any field it cannot find.
- Chinese and English pages, including photographed, skewed or curved ones. MiniCPM-OCR. PaddleOCR-VL is the other choice for Chinese-heavy files with ancient texts or rare characters.
- High-volume parsing on NVIDIA GPUs. Nemotron 3.5 Parse lets you choose one, two or three passes per page, and NVIDIA’s own table lists 13.5 pages per second for it against 2.8 for LightOnOCR-3-4B. Its card names Linux with H100 or H200 GPUs as the tested setup.
- Long PDFs where context across pages matters. Unlimited-OCR reads dozens to more than 100 pages in a single pass.
- Handwriting, checkboxes, math and many languages. Chandra OCR 2 covers 90+ languages, but read its license first.
- Layout, reading order and tables from a very small model. Surya OCR 2 does all three in one 1.4 GB model served through vLLM or llama.cpp, so it can run on a CPU or Apple Silicon. Its weights are free only below US$5 million of revenue and funding, and Datalab’s own numbers show it weak on old scans.
- Layout boxes, figure captions and chart data. LightOnOCR-3 returns them from the same model, though its model card lists English only. MiniCPM-OCR vs LightOnOCR-3 compares the two.
- Menus, posters and screenshots you want translated. HunyuanOCR translates text inside images, subject to its license.
- A phone or an older laptop. Use the 0.8B to 1.3B models. MiniCPM-OCR supports iOS and llama.cpp, and HunyuanOCR runs locally through llama.cpp.
Check the license first
Nine of the twelve current models are Apache 2.0, MIT or OpenMDW-1.1, which allow commercial use with few conditions. Three are different, and all are described in detail on their license pages:
- Chandra OCR 2 uses a modified OpenRAIL-M license. Commercial use is free only for organizations with under US$2 million in prior-year revenue and under US$2 million raised in total. It also bars using the model to compete with Datalab, and derivatives and outputs must carry the same license.
- Surya OCR 2 uses the same modified OpenRAIL-M text with higher limits: US$5 million of prior-year revenue and US$5 million raised, plus the same ban on competing with Datalab and the same share-alike rule. Its code is Apache 2.0, but the weights are not.
- HunyuanOCR uses the Tencent Hunyuan Community License. It is valid only outside the European Union, the United Kingdom and South Korea, and you may not use the model or its outputs to improve other AI models.
Two smaller points matter in practice. The full GLM-OCR pipeline uses a second model, PP-DocLayoutV3, under Apache 2.0, so two licenses apply. And the first DeepSeek-OCR is MIT while DeepSeek-OCR 2 is Apache 2.0, so check which repository you deploy. AI licenses explained covers how to read these terms.
Do not rank them by their headline scores
Almost every OCR model card quotes a score on OmniDocBench, but the numbers are not comparable. GLM-OCR reports 94.62 on version 1.5, and PaddleOCR-VL reports 96.33 on version 1.6. Both numbers come from the developers, and the two versions differ. Xiaomi-OCR-0 reports first place on Real5-OmniDocBench in its own table, Baidu created and reported the MarathonDocBench results for Unlimited-OCR, and the LightOnOCR-3 results come from LightOn’s own blog. Nemotron 3.5 Parse uses yet another benchmark, ParseBench, where NVIDIA reports 75.20 in its slowest mode, only 0.1 points above LightOnOCR-3-4B in NVIDIA’s own table.
Treat the scores as a way to build a shortlist. The Benchmark entry explains why, and the decision belongs to your own documents.
Test on your own pages
Pick 30 to 50 pages that look like your real input, including the worst scans, and run two or three candidates on them. Then check:
- Names, numbers and IDs, character by character. A wrong digit is the error you can least afford.
- Table structure, especially merged cells and cells with several lines.
- Reading order on multi-column pages, footnotes and sidebars.
- Handwriting, stamps and photographed pages, if you have them.
- Invented text. Models can add words that are not on the page, so compare a sample against the image.
- Output format. Make sure Markdown, HTML or JSON matches what your next step expects.
How to evaluate AI models yourself shows how to turn this into a repeatable test set.
Memory and setup
By our estimate, a 0.9B model needs about 2 GB of memory in BF16 and about 0.6 GB at 4-bit, and a 4B to 5B model needs about 10 to 12 GB in BF16. Page images and long outputs add to that, and you can try your own numbers in the VRAM calculator.
Setup is where these models differ most:
- PaddleOCR-VL’s default route needs the PaddlePaddle framework 3.2.1 or later, and macOS users are told to use Docker.
- LightOnOCR-3 in the 0.8B and 4B sizes needs Transformers 5.5.4 or newer.
- DeepSeek-OCR 2 and Unlimited-OCR use custom model code, so you must enable
trust_remote_codeand pin your library versions. - Nemotron 3.5 Parse also needs
trust_remote_codein Transformers, or you can use vLLM. It returns a string with control tokens and normalized coordinates, so you need NVIDIA’s postprocessing helpers to turn it into structure. - Full-page parsing with MiniCPM-OCR depends on a separate layout model, PP-DocLayoutV3, and you must start the model service yourself before its SDK works.
Most of these models run through vLLM or Transformers, and llama.cpp covers the ones with a GGUF build. Once the text is clean, the next step in a search pipeline is an embedding model, and how to choose an embedding model covers that.
Frequently asked questions
- What is the best open-source OCR model?
- There is no single best model. PaddleOCR-VL and GLM-OCR are good starting points for general pages, since both have about 0.9B parameters and permissive licenses. Their headline scores come from different versions of OmniDocBench and from the developers themselves, so test both on your own documents.
- Can I use these OCR models commercially?
- Most can. PaddleOCR-VL, MiniCPM-OCR, LightOnOCR-3, Xiaomi-OCR-0, DeepSeek-OCR 2 and NuExtract3 are Apache 2.0, and GLM-OCR and Unlimited-OCR are MIT, and Nemotron 3.5 Parse uses the permissive OpenMDW-1.1. Chandra OCR 2 limits commercial use to organizations under US$2 million in revenue and funding, Surya OCR 2 does the same below US$5 million, and HunyuanOCR is not licensed in the EU, the UK or South Korea.
- Do I need a GPU to run an OCR model?
- The models of about 1B parameters fit on a consumer GPU. By our estimate a 0.9B model needs about 2 GB of memory in BF16 and about 0.6 GB at 4-bit. For a Mac or a laptop without a large GPU, look for llama.cpp support, which MiniCPM-OCR and HunyuanOCR offer.
- Can a general model such as Qwen3.6 do OCR instead?
- It can read text in images, because Qwen3.6 accepts image input, and that may be enough for a few pages. The dedicated parsers have about 1B to 5B parameters against 27B or 35B for Qwen3.6, so they need far less memory, and they are built to write tables and formulas as Markdown.
- What does OmniDocBench measure?
- OmniDocBench scores how well a model turns PDF pages of many types into text, tables and formulas. Scores from different versions of the benchmark cannot be compared, so look at the version and at who ran the test.
Key terms in this guide
- Benchmark A benchmark is a standardized test used to compare models, such as a set of maths, coding or knowledge questions; scores are useful but easy to over-read.
- OCR and document parsing OCR turns an image of text into text; document parsing goes further and rebuilds tables, formulas, charts and reading order, usually as Markdown.
- OmniDocBench OmniDocBench is a benchmark for document parsing: it scores how well a model turns PDF pages of many types into text, tables and formulas.
- Retrieval-augmented generation (RAG) Retrieval-augmented generation lets a model answer from your own documents by retrieving the relevant passages and adding them to the prompt.
- Vision-language model (VLM) A vision-language model reads images, and often video, together with text, so you can ask questions about a picture, a screenshot or a document page.
Models mentioned in this guide
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| PaddleOCR-VL Baidu | 960M | – | Apache 2.0 (Permissive) | 0.662 GB Laptop or phone | |
| GLM-OCR Z.ai (Zhipu AI) | 900M | – | MIT (Permissive) | 0.621 GB Laptop or phone | |
| MiniCPM-OCR 1.0 OpenBMB | 1.3B | 256K | Apache 2.0 (Permissive) | 0.897 GB Laptop or phone | |
| LightOnOCR-3 LightOn | 853M – 4.54B | 256K | Apache 2.0 (Permissive) | 0.589 GB Laptop or phone | |
| Nemotron 3.5 Parse NVIDIA | 631M | – | OpenMDW-1.1 (Permissive) | 0.435 GB Laptop or phone | |
| Unlimited-OCR Baidu | 3.34B | – | MIT (Permissive) | 2.3 GB Laptop or phone | |
| DeepSeek-OCR 2 DeepSeek | 3.39B | – | Apache 2.0 (Permissive) | 2.34 GB Laptop or phone | |
| Xiaomi-OCR-0 Xiaomi | 873M | 256K | Apache 2.0 (Permissive) | 0.602 GB Laptop or phone | |
| HunyuanOCR Tencent | 1.12B | 128K | Tencent Hunyuan (Conditional) | 0.773 GB Laptop or phone | |
| Chandra OCR 2 Datalab | 5.3B | 256K | OpenRAIL-M (Datalab) (Conditional) | 3.66 GB Laptop or phone | |
| Surya OCR 2 Datalab | 690M | 256K | OpenRAIL-M (Datalab Surya) (Conditional) | 0.476 GB Laptop or phone | |
| NuExtract3 NuMind | 4.54B | 128K | Apache 2.0 (Permissive) | 3.06 GB Laptop or phone |