Unlimited-OCR on a Mac, tested
1,951 of 1,955
words found, official code on a Mac
73 more words than Tesseract
| Run | Words | Time |
|---|---|---|
| Official code, one page at a time | 1,951 | 297 s |
| Community GGUF in llama.cpp | 1,949 | 43 s |
| Official code, all 10 pages in one call | 1,949 | 355 s |
| Plain Tesseract 5.5.4 | 1,878 | 35 s |
Unlimited-OCR is a 3.3B MIT-licensed model from Baidu, built on DeepSeek-OCR. Its official code targets NVIDIA GPUs. Both ways ran on a Mac with the fixes below.
The fast way: llama.cpp
-
Download
llama-b12391-bin-macos-arm64.tar.gzfrom the llama.cpp releases on GitHub and unpack it. - Download the model and its vision file (3.9 GB in all):
curl -LO https://huggingface.co/sahilchachra/Unlimited-OCR-GGUF/resolve/main/Unlimited-OCR-Q8_0.gguf
curl -LO https://huggingface.co/sahilchachra/Unlimited-OCR-GGUF/resolve/main/mmproj-Unlimited-OCR-F16.gguf
Then read every page-*.jpg in the folder into book.txt:
for p in page-*.jpg; do
./llama-b12391/llama-mtmd-cli -m Unlimited-OCR-Q8_0.gguf --mmproj mmproj-Unlimited-OCR-F16.gguf \
--jinja --temp 0 -n 4096 -t 2 --image "$p" -p "document parsing." < /dev/null 2> /dev/null
done | sed -E 's/<\|det\|>[^<]*<\|\/det\|>//g' > book.txt
The sed line drops the layout tags before each block. The GGUF card says stock llama.cpp cannot load these files, but b12391 did.
The official code needs two changes
pip install "torch==2.15.0" torchvision "transformers==4.57.1" einops addict easydict matplotlib requests
Save uocr.py below and run python uocr.py page.jpg. The
line that moves it to the Mac GPU:
torch.Tensor.cuda = lambda self, *a, **k: self.to("mps") # the model code calls .cuda()
The full uocr.py
"""Unlimited-OCR on a Mac GPU. Run: python uocr.py page1.jpg [page2.jpg ...] > book.txt"""
import re
import sys
import torch
from transformers import AutoModel, AutoTokenizer
NAME = "baidu/Unlimited-OCR"
torch.Tensor.cuda = lambda self, *a, **k: self.to("mps") # the model code calls .cuda()
def load(name=NAME):
tok = AutoTokenizer.from_pretrained(name, trust_remote_code=True)
model = AutoModel.from_pretrained(name, trust_remote_code=True, use_safetensors=True, torch_dtype=torch.bfloat16)
return tok, model.eval().to("mps")
def clean(raw):
"""Drop the <|det|>label [box]<|/det|> markers and the <PAGE> marks, keep the text."""
raw = re.sub(r"<\|ref\|>.*?<\|/ref\|>", "", raw, flags=re.DOTALL)
raw = re.sub(r"<\|det\|>.*?<\|/det\|>", "", raw, flags=re.DOTALL)
return re.sub(r"<PAGE>", "", raw).strip()
def page(tok, model, path, out="uocr-out"):
raw = model.infer(tok, prompt="<image>document parsing.", image_file=path, output_path=out,
base_size=1024, image_size=640, crop_mode=True, max_length=32768,
no_repeat_ngram_size=35, ngram_window=128, eval_mode=True)
return clean(raw)
if __name__ == "__main__":
tok, model = load()
for p in sys.argv[1:]:
print(page(tok, model, p), flush=True)
- torch 2.15.0, not 2.10.0. The model card pins 2.10.0, which read nothing on my Mac (next section).
- The .cuda() line. The model code moves its inputs with .cuda(). That line sends them to the Mac GPU instead.
Keep save_results off. With it on, the model code runs Python eval() on text the model wrote.
torch 2.10.0 read nothing on my Mac
With the card's torch 2.10.0, page 1 came back as table tags and none of its 202 words:
</td></td></td></td></td></td></td></td></td></td></td></td></td></td></td></td></td></td></td></td></td></td></td></td>
In a small test on the Mac GPU, masked_scatter_ in torch 2.10.0 copied 1 value of 12. So the page image never reached the model.
torch 2.15.0 copied all 12.
All 10 pages in one call
The card's infer_multi reads a whole document in one request. It returns two values, the text and a token count:
pages = [f"page-{i:02d}.jpg" for i in range(1, 11)]
raw, tokens = model.infer_multi(tok, prompt="<image>Multi page parsing.", image_files=pages,
output_path="uocr-out", image_size=1024, max_length=32768,
no_repeat_ngram_size=35, ngram_window=1024)
print(clean(raw))
It wrote every page mark and found 1,949 words in 355 s. One page at a time was faster and read 2 more.
Unlimited-OCR-2 runs on a Mac too
Version 2 ships Triton kernels for NVIDIA GPUs, and pip has no Triton for a Mac. Its code uses them only on CUDA, so load() from uocr.py works:
tok, model = load("baidu/Unlimited-OCR-2")
print(clean(model.infer(tok, image_file="page.jpg", output_path="uocr-out", eval_mode=True)))
It found 1,912 words in 437 s: fewer than version 1, and slower. It also prints debug lines to the terminal.
Errors, and prompts that read nothing
| Error | Fix |
|---|---|
| Torch not compiled with CUDA enabled | Use load() from uocr.py. |
| this custom template is not supported, try using --jinja | Add --jinja, or --chat-template deepseek-ocr. |
| AttributeError: 'NoneType' object has no attribute 'convert' | A PDF path. Pass page images. |
Two other GGUF card prompts gave the wrong result:
-
Free OCR.read 176 words and left 9 of 10 pages empty. - On the one page it read, it first called the image illegible.
-
<|grounding|>Convert the document to markdown.wrote nothing on all 10 pages.
How I tested it
- Pages: the 10 scans of OCR in n8n, proofread Wikisource book pages, 1,955 true words, as 1920 px color JPEGs.
- Official code: baidu/Unlimited-OCR from Hugging Face, torch 2.15.0, transformers 4.57.1, Python 3.12, on the GPU of an Apple Silicon Mac with macOS 26.6. One page at a time with the card's single-image settings.
- llama.cpp b12391 with the Q8_0 file from sahilchachra/Unlimited-OCR-GGUF, a community conversion, not Baidu's.
- One run at a time, lowest priority, 2 CPU threads. Other jobs shared the Mac, so times are rough.
- Official code times leave out the model load (under 10 s). llama.cpp times include a model load for each page.
- Score: true words found in order, with case and punctuation.
What I did not test
- An NVIDIA GPU, vLLM, SGLang and Ollama.
- Unlimited-OCR-2 in llama.cpp, and GGUF files other than Q8_0.
- Tables, formulas, handwriting and non-English pages: I tested English book pages only.
Sources
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
Can Unlimited-OCR run on a Mac?
Yes. With torch 2.15.0 and one added line, the official code ran on a Mac GPU and found 1,951 of 1,955 words on 10 scanned book pages. A community GGUF in llama.cpp found 1,949 in 43 s.
Is Unlimited-OCR better than Tesseract?
On 10 scanned English book pages, the official code found 1,951 of 1,955 words in 297 s. Tesseract 5.5.4 found 1,878 in 35 s. In llama.cpp, the model found 1,949 in 43 s.
Why does Unlimited-OCR return only table tags on a Mac?
The model card pins torch 2.10.0. On my Mac (macOS 26.6), that version's masked_scatter_ failed on the GPU, so the page image never reached the language model. Install torch 2.15.0.
Related guides
- HunyuanOCR on a Mac: llama.cpp, 1.0 and 1.5, tested
- DeepSeek OCR on a scanned PDF: Mac setup and results, tested
- LM Studio OCR: 6 vision models tested on scanned pages
- Gemma 4 OCR: 5 sizes tested on a scanned book in Ollama
- PaddleOCR-VL on a scanned PDF: Mac setup and results, tested
- Ollama OCR on a scanned PDF: 2 models tested on a Mac
- EasyOCR vs Tesseract vs PaddleOCR vs RapidOCR, tested