OCRRouter

SmolDocling vs MiniCPM-V 4.5

Pick MiniCPM-V 4.5 for most real work: it is an Apache-2.0 8B model that runs on llama.cpp and vLLM with int4, GGUF and AWQ builds, and its authors claim strong OCRBench and OmniDocBench results. Pick SmolDocling only if you need bounding boxes or structured output (DocTags, HTML, LaTeX) from a 256M model and can accept a preview release that is English only and can loop on some pages.

SmolDoclinghuggingface/smoldoclingMiniCPM-V 4.5openbmb/minicpm-v
TypeOpen sourceVision model
Price1 credit / page2 credits / page
Per 1,000 pages$1.00$2.00
Typical speed1.0 s / page2.5 s / page
Best forVery cheap page conversionOCR on edge devices
InputsPDF, imagesimages
LanguagesMainly English30+ languages
Max pages per request500500
Tables✓ Yes✓ Yes
HandwritingNo✓ Yes
Math✓ YesNo
Bounding boxes✓ YesNo
ConfidenceNoNo
FormsNoNo
MultilingualNo✓ Yes
On OCRRouterAvailableAvailable

Prices are what you pay on OCRRouter. Speed is the provider's typical figure, not a benchmark we ran.

Which to choose

Choose SmolDocling if

  • It outputs DocTags, Markdown, HTML, JSON, LaTeX and bounding boxes, so you get location tags for text, tables, code, formulas and charts.
  • At 256M parameters it fits modest hardware, and it runs at about 0.35 seconds per page on an A100 with vLLM.
  • It publishes scores you can check: TEDS structure 0.82 on FinTabNet at 150dpi, full-page OCR F1 0.80 and OCRBench 338.

Choose MiniCPM-V 4.5 if

  • It is a general 8.7B VLM with Apache-2.0 weights and code, and it can answer questions about a document as well as read it.
  • It has int4, GGUF and AWQ builds and runs on llama.cpp and vLLM, which suits on-device OCR and local document Q&A on a single GPU.
  • It is not limited to English by anything in its published notes, while SmolDocling is English only and its authors point users to Granite-Docling for support.

Key differences

SmolDoclingMiniCPM-V 4.5
LicenceApache-2.0 per the model summary (the metadata tag says CDLA-Permissive-2.0)Apache-2.0 for weights and code
Size256M parameters8.7B parameters, built on Qwen3-8B and SigLIP2-400M
How you run itSelf-hosted, you pay for computeSelf-hosted with llama.cpp, vLLM or Transformers, you pay for compute. Ollama and SGLang support were still pending on the model card
Output formatsDocTags, Markdown, HTML, JSON, LaTeX, bounding boxesPlain text, Markdown, JSON. No bounding box or layout output
Published benchmarksTEDS 0.82 (FinTabNet), full-page OCR F1 0.80, OCRBench 338Claims lead on OCRBench and state of the art PDF parsing on OmniDocBench, but the card shows scores only as images, so numbers are hard to verify
LanguagesEnglish onlyNot published
Known failureEndless token repetition and missing closing tags; Docling issue #1515 reports loops that make generation very slow even on sparse pagesNo int4 VRAM figure published
StatusPreview release, superseded by Granite-DoclingCurrent 4.5 release

Sources

What each one is

SmolDocling

Tiny 256M vision model from Hugging Face and IBM that outputs DocTags. Predecessor of Granite Docling.

SmolDocling details

MiniCPM-V 4.5

8B vision model from OpenBMB (August 2025) built to run on devices, with strong OCR and document parsing.

MiniCPM-V 4.5 details

Try both on your own file

Run the file once with each model and compare the output.

PDF, PNG, JPG, WebP, HEIC. Up to 10 files at onceNo file?

Use one, fall back to the other

{"model": "huggingface/smoldocling","fallbacks": ["openbmb/minicpm-v"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

SmolDocling alternativesMiniCPM-V 4.5 alternativesBest OCR models in 2026