OCR in LM Studio: 6 models on 10 scanned pages
1,953 of 1,955
words GLM-OCR read in LM Studio
LM Studio runs OCR models on a Mac with no Python. Two of the six needed a fix first.
GLM-OCR read the most words
| Model | Words | Time |
|---|---|---|
| GLM-OCR | 1,953 | 69 s |
| Qwen3-VL 8B | 1,951 | 146 s |
| PaddleOCR-VL 1.6 | 1,950 | 46 s |
| Unlimited-OCR | 1,949 | 62 s |
| DeepSeek-OCR | 1,947 | 60 s |
| dots.ocr | 1,946 | 417 s |
| Tesseract 5.5.4, no LM Studio | 1,878 | 33 s |
| Our Actor's code, on our Mac | 1,879 | 38 s |
Each page went in as a 1920 px JPEG with the model's own OCR prompt, at temperature 0. Qwen3-VL has none, so it got a plain one. Times leave out the model load.
Which GGUF files we used
- GLM-OCR: ggml-org/GLM-OCR-GGUF, Q8_0, 1.4 GB.
- Qwen3-VL 8B: Qwen/Qwen3-VL-8B-Instruct-GGUF, Q4_K_M, 6.2 GB.
- PaddleOCR-VL 1.6: PaddlePaddle/PaddleOCR-VL-1.6-GGUF, 1.8 GB.
- Unlimited-OCR: sahilchachra/Unlimited-OCR-GGUF, Q8_0, 3.9 GB.
- DeepSeek-OCR, not version 2: ggml-org/DeepSeek-OCR-GGUF, Q8_0, 3.6 GB.
- dots.ocr: ggml-org/dots.ocr-GGUF, Q8_0, 3.2 GB.
- Sizes count the model file and its vision (mmproj) file.
Two models need a fix first
- PaddleOCR-VL: no images. LM Studio 0.4.26 answered "ocr-paddle does not support image inputs." The vision file's name ends with mmproj. Renamed to start with mmproj, it read images.
cd ~/.lmstudio/models/PaddlePaddle/PaddleOCR-VL-1.6-GGUF
mv PaddleOCR-VL-1.6-GGUF-mmproj.gguf mmproj-PaddleOCR-VL-1.6-GGUF.gguf
- Unlimited-OCR: box numbers. LM Studio hides the model's layout tags, so a label and four box numbers stick to the first word of each block. It found 1,910 words as returned. Our script's sed line strips them.
- New files showed after a restart. After lms import, LM Studio listed our new models only once it restarted.
Install once, then load GLM-OCR
With LM Studio on the Mac, get poppler and the model:
brew install poppler
~/.lmstudio/bin/lms get https://huggingface.co/ggml-org/GLM-OCR-GGUF -y
~/.lmstudio/bin/lms load glm-ocr --identifier ocr
On our Mac, lms get picked the Q8_0 files we tested. LM Studio's server refused our PDF, so the script turns each page into a JPEG first.
One script: scanned PDF in, text out
Save this as lmocr.sh:
pdftoppm -r 300 -jpeg -jpegopt quality=95 "$1" page
for p in page-*.jpg; do
{ printf '{"model":"ocr","temperature":0,"max_tokens":4096,"messages":[{"role":"user","content":[{"type":"image_url","image_url":{"url":"data:image/jpeg;base64,'
base64 -i "$p" | tr -d '\n'
printf '"}},{"type":"text","text":"%s"}]}]}' "$2"
} > req.json
curl -s localhost:1234/v1/chat/completions -H 'Content-Type: application/json' -d @req.json > out.json
plutil -extract choices.0.message.content raw -o - out.json
echo
done | sed -E 's/^[a-z_]+ ?\[\[?[0-9]+, [0-9]+, [0-9]+, [0-9]+\]\]?//'
Then run it in the folder with scan.pdf:
sh lmocr.sh scan.pdf "Text Recognition:" > book.txt
On our scan it found 1,949 words in 81 s, model load included. The words it lost came back as italics, like *The Jolly Ruffian*.
The prompt for each model
- GLM-OCR: Text Recognition:
- Qwen3-VL: Read all the text in the image.
- PaddleOCR-VL: OCR:
- Unlimited-OCR: document parsing.
- DeepSeek-OCR: <|grounding|>Convert the document to markdown.
- dots.ocr: Extract the text content from this image.
Temperature 0 gave the same text twice
| Setting | Words |
|---|---|
| Temperature 0, first run | 1,953 |
| Temperature 0, second run | 1,953, same text |
| LM Studio defaults | 1,948 |
Every page fit LM Studio's default context of 8,192 tokens, in every model. dots.ocr came closest: 7,935 tokens for its biggest page.
Same 10 pages, one scorer
The 10 pages are proofread scans from Wikisource, 1,955 words in all, the same pages as our OCR roundup. The score is true words found, in order, with exact case and punctuation.
How we ran it
- LM Studio 0.4.26 with its llama.cpp engine 2.73.0, driven by its lms tool and its server.
- Each GGUF file matched the SHA-256 on Hugging Face.
- One Apple Silicon Mac, one model at a time, at low priority, so times are rough.
- Before scoring, we removed layout labels and box numbers. That moved DeepSeek-OCR by 1 word.
- Licenses: MIT or Apache-2.0 for the models. dots.ocr adds its own agreement, and our public domain scan fits it.
Scanned pages to text, no install
Run it yourself. Our Apify Actor reads PDF, PNG, JPEG, WEBP and TIFF for $0.006 a page, with nothing to install. 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 OCR model in LM Studio?
On our scan, GLM-OCR: 1,953 of 1,955 words in 69 s. Qwen3-VL 8B came second and took about twice as long.
Can LM Studio OCR a PDF?
Not through its server: it sent back an error for our PDF. Turn each page into a JPEG with pdftoppm first, as the script above does.
Why does PaddleOCR-VL say it does not support images?
In LM Studio 0.4.26 it read images only after we renamed its vision file to start with mmproj, as shown above, and restarted LM Studio.
Does DeepSeek-OCR work in LM Studio?
Yes. The ggml-org GGUF of the original DeepSeek-OCR loaded in LM Studio 0.4.26 and found 1,947 of 1,955 words.
Can I use LM Studio for OCR at work?
Its terms of Aug 23, 2026 allow personal and internal business use. They bar running it as a service for other people. Each model has its own license too.
Related guides
- GLM-OCR on a scanned PDF: Mac setup and results, tested
- Ollama OCR on a scanned PDF: 2 models tested on a Mac
- Baidu Unlimited OCR on a Mac: llama.cpp and Python, tested
- Qwen3-VL OCR on a scanned PDF: tested in Ollama on a Mac
- Best open source OCR in 2026: 28 engines on 10 book pages
- MLX OCR on a Mac: 6 OCR models in mlx-vlm, tested