SmolDocling tested on a scanned book
1,891
of 1,955 words on a 10-page scan
Its successor won by 5 words
Same 10 proofread book pages, same Mac, same morning.
| Tool | Words | Time |
|---|---|---|
| Granite-Docling 258M | 1,896 | 57 s |
| SmolDocling 256M | 1,891 | 53 s |
| Our Actor's code | 1,879 | 35 s |
| Tesseract 5.5.4 | 1,878 | 31 s |
Two SmolDocling runs gave the same text. The first run took 49 s longer, 43 s of it to fetch the 518 MB model.
It keeps line-end hyphens
-
SmolDocling left 6 words split at the line end, like
gath- ered. Granite-Docling left 2. -
Both put spaces inside quote marks:
" Who's. With quote marks ignored, they almost tie: Granite-Docling 1,933, SmolDocling 1,932. -
SmolDocling also read mending as
meningand yourn asyour.
Run it on a Mac
pip install docling==2.150.0 mlx-vlm==0.7.12
docling --to md --pipeline vlm --vlm-model smoldocling book.pdf
Use an arm64 Python. With an Intel Python, pip finds no mlx wheel and stops.
The SmolDocling preset sets no output cap, so docling's default applies: 4,096 tokens of output a page. Granite-Docling's preset allows 8,192.
Two license lines, both open
- The Hub tag and card header say CDLA-Permissive-2.0.
- The card's Model Summary says Apache 2.0.
- Both allow use, and neither limits what you do with the output.
I read the CDLA-Permissive-2.0 in full. This is not legal advice.
Pick Granite-Docling for new work
The SmolDocling card calls Granite-Docling its successor: it "will now receive updates and support". The Granite-Docling card says it tends to avoid infinite loops more effectively.
How we ran it
-
A Mac with Apple M5 Max. docling 2.150.0 and mlx-vlm 0.7.12, the
presets
smoldoclingandgranite_docling,--num-threads 2, one run at a time at low priority. - Models: SmolDocling-256M-preview-mlx-bf16 (revision 4439622) and granite-docling-258M-mlx (revision e9939db), both picked by docling on a Mac.
- The scan: 10 proofread pages from two public domain books on Wikisource, 1,955 words, the same PDF as our earlier tests.
- Score: true words found in order, after image links and comments are removed. Extra text, like running heads, costs nothing.
- Times have the models already on disk. Tesseract and the Actor's code ran from 05:17 the same day.
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.
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 SmolDocling?
A 256M vision model from IBM Research and Hugging Face. It reads a page image and writes DocTags, which docling turns into Markdown, HTML or JSON. Its card lists English only.
Does SmolDocling run on a Mac?
Yes. On Apple silicon, docling picks the MLX copy of the model. It read our 10 pages in 53 s on the Mac GPU. Use an arm64 Python: mlx has no Intel Mac wheels.
SmolDocling or Granite-Docling?
Granite-Docling. The SmolDocling card names it the successor that gets updates and support. On our scan it found 5 more words in about the same time.
Can I use SmolDocling in a business?
Its Hub tag says CDLA-Permissive-2.0 and its card text says Apache 2.0. Both allow use and put no limits on what you do with the output. This is not legal advice.
Can SmolDocling read a PDF?
Yes, through docling: it draws each PDF page as an image, the model reads it, and docling writes one Markdown file for the whole PDF.