InternVL3.5 vs MiniCPM-V 4.5
Pick InternVL3.5 if you want one model family you can size to your hardware, from 1B up to 241B-A28B, for document question answering in a self-hosted pipeline. Pick MiniCPM-V 4.5 if you want a single 8.7B model that runs locally or on edge devices through llama.cpp, vLLM, int4, GGUF or AWQ builds. Both are Apache-2.0, self-hosted and return no bounding boxes, so the real split is size flexibility versus a fixed small model with quantized builds.
InternVL3.5opengvlab/internvl | MiniCPM-V 4.5openbmb/minicpm-v | |
|---|---|---|
| Type | Vision model | Vision model |
| Price | 2 credits / page | 2 credits / page |
| Per 1,000 pages | $2.00 | $2.00 |
| Typical speed | 3.0 s / page | 2.5 s / page |
| Best for | Self-hosted document understanding | OCR on edge devices |
| Inputs | images | images |
| Languages | Multilingual | 30+ languages |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | ✓ Yes |
| Math | ✓ Yes | No |
| Bounding boxes | No | No |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | ✓ Yes | ✓ 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 InternVL3.5 if
- You can choose from nine sizes between 1B and 241B, so you can trade accuracy for cost on your own hardware.
- The 20B-A4B and 30B-A3B MoE variants activate few parameters per token, and models up to 30B fit on a single A100.
- It suits teams that want one model family for document question answering across several deployment sizes.
Choose MiniCPM-V 4.5 if
- It ships as int4, GGUF and AWQ builds on llama.cpp and vLLM, which fits on-device or edge OCR.
- OpenBMB claims leading OCRBench results, above GPT-4o-latest and Gemini 2.5, and state of the art PDF parsing among general MLLMs on OmniDocBench.
- A single 8.7B model is enough for local document Q&A on one GPU, with no need to pick a size.
Key differences
| InternVL3.5 | MiniCPM-V 4.5 | |
|---|---|---|
| Licence | Apache-2.0 | Apache-2.0 weights and code |
| Model sizes | Nine sizes, 1B to 241B-A28B (8B variant is 8.5B) | One size, 8.7B, built on Qwen3-8B and SigLIP2-400M |
| How you run it | Transformers, LMDeploy or vLLM | llama.cpp, vLLM, Ollama or Transformers (Ollama and SGLang support still pending official releases on the model card) |
| Hardware | Up to 30B on a single A100; 38B needs two A100s; 241B-A28B needs eight A100s | No int4 VRAM figure published |
| Benchmarks | OCR scores published only as chart images | Scores given only as images; OpenBMB claims leading OCRBench and OmniDocBench PDF parsing results |
| Output formats | Plain text, Markdown, JSON | Plain text, Markdown, JSON |
| Coordinates and layout | No bounding boxes or layout output by default | No bounding box or layout output |
Sources
- InternVL3.5 8B model card (Hugging Face)
- InternVL GitHub
- MiniCPM-V 4.5 model card (Hugging Face)
- MiniCPM-V 4.5 paper
What each one is
InternVL3.5
Shanghai AI Lab's open vision-language family (1B to 241B, August 2025), strong on OCR and document benchmarks.
InternVL3.5 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": "opengvlab/internvl","fallbacks": ["openbmb/minicpm-v"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
InternVL3.5 alternativesMiniCPM-V 4.5 alternativesBest OCR models in 2026