Dolphin OCR tested on English and Chinese scans
1,923
of 1,955 words on a 10-page scan
English: most misses are a space
Dolphin-1.5 read 10 proofread book pages on a Mac with an M5 Max.
| Way | Words | Time |
|---|---|---|
| Page JPEGs, Mac GPU | 1,923 | 278 s |
| Page JPEGs, CPU as shipped | 1,923 | 449 s |
| The PDF, Mac GPU | 1,921 | 275 s |
| Page JPEGs, then tidy.py | 1,947 | 278 s |
Of its 31 misses (32 words), 25 are only a space:
" Who's for "Who's, and
it ? for it?. Our tidy.py takes those spaces
out.
It misread two words: stayed for staved, and oldfashioned for old-fashioned.
tidy.py on 45 pages it never saw
I wrote tidy.py after I saw those misses. So I ran Dolphin on 45 more pages from 9 other Wikisource books.
| Run | Words |
|---|---|
| Dolphin-1.5 | 12,298 |
| Dolphin-1.5, then tidy.py | 12,393 |
| Tesseract 5.5.4 | 11,975 |
tidy.py still helped, but it cost 2 words on one page. One old book prints the long s (ſ), and Dolphin wrote fave for save.
Chinese: 96.3% against Tesseract's 88.4%
| Pages | Dolphin-1.5 | Tesseract |
|---|---|---|
| Simplified, 36 pages | 96.3% | 88.4% |
| Simplified, less one bad page | 98.4% | 90.3% |
| Traditional, vertical, 10 pages | 10.4% | 52.9% |
The pages are proofread Chinese Wikisource scans from our Tesseract Chinese test. Tesseract used chi_sim, and chi_tra_vert for vertical pages.
The bad page is from a gazette. Its proofread text holds 679 characters, but the scan shows about half. Both tools wrote 362.
Vertical pages are a trap. Dolphin finished only 4 of 10 within 4 minutes, and on some it read across the columns. The rest count as zero.
On the 4 it finished, it found 29.1% and Tesseract 61.6%.
Run it on a Mac
git clone https://github.com/bytedance/Dolphin.git
cd Dolphin
git checkout fa8d6be4
python3.12 -m venv venv
. venv/bin/activate
pip install torch==2.6.0 torchvision==0.21.0 transformers==4.47.0 \
accelerate==1.6.0 "numpy<2" opencv-python-headless pymupdf timm \
omegaconf "huggingface_hub<0.30"
huggingface-cli download ByteDance/Dolphin-1.5 --local-dir hf_model
python demo_page.py --model_path hf_model --input_path pages --save_dir out
pages is a folder of page images. A PDF works too, but
the script draws each page only 896 px on its long side.
Then join the pages and fix the spaces with tidy.py, standard library only:
python3 tidy.py out/markdown/*.md > book.txt
On a Mac the script uses the CPU: it looks for CUDA only. Change
"cpu" to "mps" in its
device line to use the Mac GPU.
Use Dolphin-1.5, not v2, in a business
| Part | License |
|---|---|
| Dolphin-1.5 weights | MIT |
| GitHub branch v1.5, or commit fa8d6be4 | MIT |
| GitHub master branch since Dec 17, 2025 | Qwen Research License |
| Dolphin-v2 weights | Qwen Research License |
The Qwen Research License allows research and evaluation only. For other use, it says to ask Alibaba Cloud for a license.
This page tests only MIT parts: the 1.5 weights and the code at commit fa8d6be4. This is not legal advice: read both files yourself.
Ahead of Tesseract, behind the big models
| Tool | Words | Time |
|---|---|---|
| MiniCPM-OCR 1.0, llama.cpp | 1,954 | 16 s |
| olmOCR-2 7B | 1,954 | 77 s |
| LightOnOCR-2 1B | 1,954 | 72 s |
| Dolphin-1.5, then tidy.py | 1,947 | 278 s |
| PaddleOCR-VL 1.6 4-bit in MLX | 1,942 | 20 s |
| Dolphin-1.5 | 1,923 | 278 s |
| Our Actor's code | 1,879 | 38 s |
| Tesseract 5.5.4 | 1,878 | 34 s |
| GOT-OCR 2.0 | 1,591 | 178 s |
Times are rough: the other rows ran on earlier days, and Dolphin ran at low priority on a busy Mac.
How we ran it
- A Mac with Apple M5 Max. Dolphin-1.5 from Hugging Face, commit 0fe42a93: 398 million parameters, 796 MB in float16.
- Code: bytedance/Dolphin at commit fa8d6be4, unchanged. Python 3.12.12, torch 2.6.0 (the repo pins 2.1.0, which has no Python 3.12 wheel), transformers 4.47.0.
- Every run at low priority (nice and background QoS), 3 threads, one at a time, while other jobs ran on the Mac.
- English: 10 proofread pages from two public domain books on Wikisource, 1,955 words. Holdout: 45 pages from 9 more books, each with the same word count as in our earlier fetch.
- Chinese: the 72 pages of our Tesseract Chinese test, fetched again at the same proofread revisions. All 72 matched.
- Scores: true words found in order, or Chinese characters found in order. Extra text, like running heads, costs nothing.
- Vertical pages: the full run died twice on page 1. The second time its memory footprint reached 165 GB. So I ran the first page of each of the 10 books alone, stopped after 240 s. Tesseract read the same pages.
- Tesseract on Chinese ran again today and matched its Nov 8 score on both sets.
Scanned pages to text, no install
Run it yourself. Our Apify Actor turns a scanned PDF into Markdown or text. $0.006 a scanned page. Its code found 1,879 of these 1,955 words. For Chinese it uses the chi_sim file of the table above.
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 Dolphin OCR?
A document parsing model from ByteDance, described in a paper in Findings of ACL 2025. Dolphin-1.5 has 398 million parameters: a Swin image encoder and an MBart text decoder. Its card lists Chinese and English.
Does Dolphin run on a Mac?
Yes. The script runs on the CPU, as shipped. With one line changed it ran on the Mac GPU and gave the same text.
Can I use Dolphin in a business?
Dolphin-1.5 weights are MIT, and so is the GitHub code in branch v1.5 or at commit fa8d6be4. The master branch since Dec 17, 2025, and Dolphin-v2 use the Qwen Research License: research and evaluation only.
Does Dolphin read Chinese?
Yes, in lines across the page. On 36 simplified Chinese scans it found 96.3% of the characters in order. Tesseract found 88.4%. On 10 vertical pages it found 10.4%, and Tesseract 52.9%. It finished only 4 of them in 4 minutes.
How do I fix the spaces around quote marks?
Run tidy.py on the Markdown files. It takes out the space after an opening quote and before a closing quote or a question mark. It lifted our scan from 1,923 to 1,947 words.