Granite Docling vs MiniCPM-V 4.5
Pick Granite Docling for English PDFs with tables, code and equations, especially inside a Docling pipeline on CPU or a small GPU. It is only 258M parameters and outputs DocTags with bounding boxes. Pick MiniCPM-V 4.5 when you need a general 8.7B vision language model for on-device OCR or local document Q&A, and do not need coordinates.
Granite Doclingibm/granite-docling | MiniCPM-V 4.5openbmb/minicpm-v | |
|---|---|---|
| Type | Open source | Vision model |
| Price | 1 credit / page | 2 credits / page |
| Per 1,000 pages | $1.00 | $2.00 |
| Typical speed | 1.5 s / page | 2.5 s / page |
| Best for | Small-model document conversion | OCR on edge devices |
| Inputs | PDF, images | images |
| Languages | English, experimental CJK and Arabic | 30+ languages |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | ✓ Yes |
| Math | ✓ Yes | No |
| Bounding boxes | ✓ Yes | No |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | No | ✓ Yes |
| On OCRRouter | Available | Available |
Prices are what you pay on OCRRouter. Speed is the provider's typical figure, not a benchmark we ran.
Which to choose
Choose Granite Docling if
- You run a Docling pipeline: it outputs DocTags, Markdown, HTML, LaTeX and bounding boxes.
- Your documents are English PDFs with tables, code and equations: table structure TEDS is 0.97 on FinTabNet, code F1 is 0.988 and equation F1 is 0.968, per the model card.
- You want low-cost conversion on CPU or a small GPU, since the model has only 258M parameters.
Choose MiniCPM-V 4.5 if
- You need a general vision language model that also answers questions about a document, run locally on a single GPU.
- You want quantized deployment on edge devices: int4, GGUF and AWQ builds run on llama.cpp and vLLM.
- Your documents are not English-first, and you want to avoid Granite Docling's experimental status for Japanese, Arabic and Chinese.
Key differences
| Granite Docling | MiniCPM-V 4.5 | |
|---|---|---|
| Size | 258M parameters | 8.7B parameters, built on Qwen3-8B and SigLIP2-400M |
| Licence | Apache-2.0 | Apache-2.0 weights and code |
| How you run it | Self-hosted, typically inside Docling | Self-hosted with llama.cpp, vLLM, Ollama or Transformers; Ollama and SGLang support were still pending on the model card |
| Output formats | DocTags, Markdown, HTML, LaTeX, bounding boxes | Plain text, Markdown, JSON; no bounding box or layout output |
| Published benchmarks | Table TEDS 0.97 (FinTabNet), full-page OCR F1 0.84, equation F1 0.968, OCRBench 500, all from the authors | Not published as numbers; the model card shows scores only as images. OpenBMB claims leading OCRBench results and state of the art PDF parsing among general MLLMs on OmniDocBench |
| Languages | English primary; Japanese, Arabic and Chinese marked experimental | Not published |
| Known failure modes | Repetitive loops on page headings and invoices; exclamation-mark output on GPUs without bfloat16 such as T4 unless run in float32; vLLM wte attribute error needs the untied revision; Ollama FP16 overflow hallucinations on CPU | No int4 VRAM figure published; benchmark numbers hard to verify |
Sources
- Granite-Docling model card
- HF discussion #24 repetition
- Ollama issue #12355
- Hugging Face model card
- GitHub repo
What each one is
Granite Docling
Compact IBM vision model that converts pages to DocTags for the Docling pipeline.
Granite Docling detailsMiniCPM-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 detailsTry both on your own file
Run the file once with each model and compare the output.
Use one, fall back to the other
{"model": "ibm/granite-docling","fallbacks": ["openbmb/minicpm-v"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Granite Docling alternativesMiniCPM-V 4.5 alternativesBest OCR models in 2026