olmOCR vs MiniCPM-V 4.5
Pick olmOCR for bulk PDF to Markdown on a GPU you control, especially academic papers with equations and pages that mix print and handwriting. Pick MiniCPM-V 4.5 for local or on-device OCR on quantized builds, or for document Q&A on a single GPU. The main split is that olmOCR is a purpose-built converter with a published benchmark score, while MiniCPM-V 4.5 is a general vision model with no published layout output.
olmOCRallenai/olmocr | MiniCPM-V 4.5openbmb/minicpm-v | |
|---|---|---|
| Type | Open source | Vision model |
| Price | 2 credits / page | 2 credits / page |
| Per 1,000 pages | $2.00 | $2.00 |
| Typical speed | 1.5 s / page | 2.5 s / page |
| Best for | Bulk PDF to text for datasets | OCR on edge devices |
| Inputs | PDF, images | images |
| Languages | Mainly English | 30+ languages |
| Max pages per request | 1,000 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | ✓ Yes | ✓ Yes |
| Math | ✓ Yes | No |
| Bounding boxes | No | 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 olmOCR if
- It scores 82.4 on olmOCR-Bench, above Marker (76.1), MinerU 2.5.4 (75.2) and Mistral OCR (72.0) in the same table.
- Self-hosting costs under $200 per million pages, per Ai2, and a hosted olmOCR-2-7B endpoint on DeepInfra costs $0.09 input and $0.19 output per million tokens.
- It handles equations, tables and handwriting, and strips headers and footers in natural reading order.
Choose MiniCPM-V 4.5 if
- It ships int4, GGUF and AWQ builds and runs on llama.cpp and vLLM, which suits on-device and edge use.
- It is a general vision language model, so you can ask questions about a document locally on one GPU instead of only converting it.
- It outputs plain text, Markdown or JSON, and both weights and code are Apache-2.0.
Key differences
| olmOCR | MiniCPM-V 4.5 | |
|---|---|---|
| Model size | 7B | 8.7B (Qwen3-8B with SigLIP2-400M) |
| How you run it | Self-host with vLLM on an NVIDIA GPU, or use third-party hosted endpoints | Self-host only, with llama.cpp, vLLM, Ollama or Transformers |
| Licence | Apache-2.0 (toolkit) | Apache-2.0 (weights and code) |
| Hardware | NVIDIA GPU with at least 12 GB VRAM and 30 GB free disk | No int4 VRAM figure published |
| Price | Under $200 per million pages self-hosted; DeepInfra $0.09 / $0.19 per million tokens | Free to self-host, you pay for compute |
| Benchmark | 82.4 on olmOCR-Bench (v0.4.0), from the authors | Not published as text. The model card gives scores only as images, and OpenBMB claims leading OCRBench results |
| Known failure | Out of memory on a 12 GB GPU, can get stuck on the last few pages, and image bounding box coordinates are unreliable | No bounding box or layout output; Ollama and SGLang support were still pending on the model card |
Sources
What each one is
olmOCR
Open-source vision model from Ai2 (often searched as olm ocr) that turns PDFs into clean, reading-order text. Latest release: olmOCR 2.
olmOCR 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": "allenai/olmocr","fallbacks": ["openbmb/minicpm-v"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
olmOCR alternativesMiniCPM-V 4.5 alternativesBest OCR models in 2026