Best open source OCR, tested on 10 book pages
1,954 of 1,955
words, olmOCR-2 and LightOnOCR-2 1B
The top 10 are all vision models
| Engine | Words | Time |
|---|---|---|
| olmOCR-2 7B | 1,954 | 77 s |
| LightOnOCR-2 1B | 1,954 | 72 s |
| LightOnOCR-3 | 1,953 | 130 s |
| Unlimited-OCR | 1,951 | 297 s |
| Qwen3-VL 8B | 1,951 | 139 s |
| GLM-OCR | 1,950 | 118 s |
| Gemma 4 e4b (Ollama) | 1,950 | 32 s |
| MinerU 4.2.1 | 1,949 | 41 s |
| DeepSeek-OCR 2 | 1,948 | 83 s |
| PaddleOCR-VL 1.6 | 1,947 | 73 s |
The best classic engine, PaddleOCR, found 1,919 on a CPU. Tesseract found 1,878 in 31 s.
Times are rough. Most include the model load. Some ran on the Mac GPU, some on the CPU, on different days.
Pick by what you have
| You have | Pick | Words |
|---|---|---|
| A GPU (I tested a Mac) | LightOnOCR-2 1B | 1,954 |
| Ollama and one command | Gemma 4 e4b | 1,950 |
| A CPU and Python | PaddleOCR 3.7 | 1,919 |
| A CPU and half a minute | Tesseract 5.5.4 | 1,878 |
| Node.js | Scribe.js (AGPL-3.0) | 1,895 |
| A scanned PDF to search | OCRmyPDF (Tesseract inside) | 1,855 |
LightOnOCR-2 has 1B parameters, a seventh of olmOCR-2's 7B. Each link goes to its own test, with the command and the traps.
All 28 engines, same pages
16 vision language models and 12 classic engines. Each row is the run its own page leads with. tesseract.js shares the Scribe.js page, and Tesseract is tonight's rerun.
16 vision language models
| Engine | Words | Time |
|---|---|---|
| olmOCR-2 7B | 1,954 | 77 s |
| LightOnOCR-2 1B | 1,954 | 72 s |
| LightOnOCR-3 | 1,953 | 130 s |
| Unlimited-OCR | 1,951 | 297 s |
| Qwen3-VL 8B | 1,951 | 139 s |
| GLM-OCR | 1,950 | 118 s |
| Gemma 4 e4b (Ollama) | 1,950 | 32 s |
| MinerU 4.2.1 | 1,949 | 41 s |
| DeepSeek-OCR 2 | 1,948 | 83 s |
| PaddleOCR-VL 1.6 | 1,947 | 73 s |
| dots.ocr | 1,946 | 168 s |
| GOT-OCR 2.0 | 1,936 | 134 s |
| Surya | 1,926 | 56 s |
| Chandra 2 | 1,911 | 169 s |
| Granite-Docling 258M | 1,897 | 54 s |
| Florence-2 | 1,887 | 158 s |
12 classic engines
| Engine | Words | Time |
|---|---|---|
| PaddleOCR 3.7 (PP-OCRv6) | 1,919 | 319 s |
| Scribe.js 0.18.2 | 1,895 | 15 s |
| RapidOCR 3.10.0 | 1,894 | 55 s |
| Tesseract 5.5.4 | 1,878 | 31 s |
| tesseract.js 7.0.0 | 1,876 | 18 s |
| Kraken 7.1.1 | 1,875 | 380 s |
| docTR | 1,869 | 13 s |
| Pix2Text 1.1.7 | 1,810 | 99 s |
| ocrs 0.14.0 (Rust) | 1,743 | 35 s |
| EasyOCR 1.7.2 | 1,658 | 58 s |
| OpenCV 5.0.0 (EAST + CRNN) | 1,533 | 283 s |
| keras-ocr 0.9.3 | 1,334 | 870 s |
Settings behind four rows
-
PaddleOCR: the page's
--fastsetup. The default read 1,908. - GOT-OCR 2.0: format mode. Plain mode read 1,591.
- Qwen3-VL: temperature 0. The Ollama default read 1,950.
- PaddleOCR-VL: the GGUF in llama-server, not the CPU path of the model card.
Six licenses with a catch
- olmOCR-2: Apache-2.0. The card says research and educational use, but Ai2's guidelines do not ban business use.
- MinerU: Apache-2.0 base, with its own terms above 100 million monthly users or $20 million monthly revenue.
- Surya: Apache-2.0 code. The weights are free for research, personal use and startups under $5M.
- Chandra 2: Apache-2.0 code. The weights are free under $2M in revenue and funding, and not for businesses that compete with Datalab.
- Scribe.js: AGPL-3.0. Read it before you ship it in a product.
- Pix2Text: MIT, but pip installs doclayout-yolo and ultralytics, both AGPL-3.0.
dots.ocr has its own model agreement, which allows business use. The other 21 are Apache-2.0 or MIT, two with Creative Commons models.
Same pages, same scorer
Every engine read the same 10 proofread book pages from Wikisource, 1,955 words in all. One script scores every run: true words found, in order, with exact case and punctuation (curly quotes and dashes count as plain ones).
How I know the old runs match
- Most runs read a PDF of the 10 page images from Nov 10. A reboot on Nov 14 deleted it, so the later runs used a rebuilt copy.
- Both PDFs keep the JPEG bytes as they are, at a 300 dpi page size. The image hashes from Nov 11 match today's.
- Tonight RapidOCR read today's PDF and gave the same text, byte for byte, as its Nov 13 run on the Nov 10 PDF. Three runs from earlier today also matched byte for byte:
| Rerun | Then | Tonight |
|---|---|---|
| RapidOCR, from the Nov 10 PDF | 1,894 | 1,894 |
| Unlimited-OCR in llama.cpp | 1,949 | 1,949 |
| Tesseract 5.5.4 | 1,878 | 1,878 |
| The Actor's code | 1,879 | 1,879 |
The vision models, docTR, Kraken and EasyOCR ran on the GPU of one Apple Silicon Mac. The other classic engines ran on the CPU, most on 2 threads. Every run had low priority.
What 10 pages cannot tell you
- Extra words are not counted. The score does not punish running heads or a line that is not on the page.
- One kind of page. Clean printed English. No handwriting, tables, forms, phone photos or other languages.
- Near the top, it is a tie. The top 11 are 8 words apart, so pick by license, size and speed.
Scanned pages to text, no GPU
Run it yourself. Our Apify Actor reads PDF, PNG, JPEG, WEBP and TIFF for $0.006 a page, with no model files. Its code found 1,879 of these 1,955 words.
Or send us the scan. Tack, our AI, does the job for $29, up to 50 pages. First page free.
Questions people ask
What is the best open source OCR in 2026?
On 10 scanned English book pages, olmOCR-2 7B and LightOnOCR-2 1B found 1,954 of 1,955 words. Both are Apache-2.0 vision models. olmOCR-2's card adds a research and education note. On a CPU, PaddleOCR found 1,919.
What is the best open source OCR for PDF files?
To make a scanned PDF searchable, OCRmyPDF found 1,855 of 1,955 words with Tesseract inside. To get the text out, the olmOCR toolkit takes the PDF itself and found 1,954.
Is Tesseract still good enough?
For clean printed pages, often yes. Tesseract 5.5.4 found 1,878 of 1,955 words in 31 s on a CPU. The two best vision models found 76 more words.
What is the best open source OCR for handwriting?
This test has no handwriting. Our TrOCR and Tesseract handwriting pages test 100 handwritten notes.
Related guides
- Free ABBYY FineReader alternatives for Mac and Linux, tested
- LM Studio OCR: 6 vision models tested on scanned pages
- olmOCR on a Mac: the toolkit with llama.cpp, tested
- LightOnOCR on a scanned PDF: Mac setup and results, tested
- PaddleOCR on a Mac: CPU setup and results, tested
- EasyOCR vs Tesseract vs PaddleOCR vs RapidOCR, tested
- How to measure OCR accuracy: CER and WER in Python, tested