MiniCPM-V Multimodal Series
MiniCPM-V is a series of multimodal vision-language models built for on-device deployment, supporting image and video understanding, with sizes ranging from 1B to 35B parameters. The series also includes MiniCPM-OCR 1.0 for document parsing and MiniCPM-OCR Jev for OCR quality assessment.
Series Overview
| Model | Parameters | Architecture | VRAM | Highlights | HuggingFace |
|---|---|---|---|---|---|
| MiniCPM-OCR 1.0 | 1.3B | SigLIP + Qwen3.5, trained from MiniCPM-V 4.6 | — | Chinese/English document parsing, recognition and KIE | Download |
| MiniCPM-OCR Jev | 2B | Qwen3.5-2B + binary classification head | ≥12 GB (local tutorial) | OCR region quality assessment, candidate selection | Download |
| MiniCPM-V 4.7 35B-A3B | 35.2B total · ~3.5B active | SigLIP canvas M-RoPE + Qwen3.5-MoE | — | First MoE in the series, runs smoothly on device | Download |
| MiniCPM-V 4.7 1B | 1.55B | SigLIP canvas M-RoPE + Qwen3.5-0.8B | — | 256K context, on-device friendly | Download |
| MiniCPM-V 4.6 | 1.3B | SigLIP2 + Qwen3.5-0.8B | 4 GB | Early ViT compression, on-device deployment | Download |
| MiniCPM-V 4.6 Thinking | 1.3B | SigLIP2 + Qwen3.5-0.8B | 4 GB | Deep reasoning variant | Download |
| MiniCPM-V 4.5 | 8.7B | SigLIP2 + Qwen3-8B | ~28 GB | Surpasses GPT-4o-latest, 30+ languages | Download |
| MiniCPM-V 4.0 | 4.1B | SigLIP2 + MiniCPM4-3B | — | Surpasses GPT-4.1-mini | Download |
Recommended Choices
| Scenario | Recommended model |
|---|---|
| Chinese/English PDF, scan and photographed document parsing | MiniCPM-OCR 1.0 |
| OCR result screening and multi-model candidate selection | MiniCPM-OCR Jev |
| Mobile / ultra-lightweight deployment | MiniCPM-V 4.7 1B |
| High-throughput serving / peak performance | MiniCPM-V 4.7 35B-A3B |
| High-accuracy image/video understanding | MiniCPM-V 4.5 |
| Deep reasoning (math/code) | MiniCPM-V 4.7 (enable_thinking) or MiniCPM-V 4.6 Thinking |
| Balanced performance | MiniCPM-V 4.0 |
On-device Deployment
MiniCPM-V 4.7 and 4.6 can run fully offline on mobile devices (iOS / Android / HarmonyOS). Download: MiniCPM-V-Apps
Related Resources
| Topic | Link |
|---|---|
| Code repository and release notes | OpenBMB/MiniCPM-V / OpenBMB/MiniCPM-OCR |
| MiniCPM-V 4.7 35B-A3B model weights | Hugging Face / ModelScope |
| MiniCPM-V 4.7 1B model weights | Hugging Face / ModelScope |
| MiniCPM-OCR 1.0 model weights | Hugging Face / ModelScope |
| MiniCPM-OCR Jev model weights | Hugging Face / ModelScope |
| Model documentation | MiniCPM-OCR / MiniCPM-V 4.7 / MiniCPM-V 4.6 / MiniCPM-V 4.5 |
| Online demo | MiniCPM-V 4.7 Demo |
| On-device apps and downloads | MiniCPM-V-Apps downloads |
| API access | API guide |
| Technical reports and architecture papers | MiniCPM-V 4.5 / LLaVA-UHD v4 |
| Multimodal technical community group | Join the MiniCPM-V community |
Developer Showcase
View all →MiniCPM Agent Gateway: Edge-Cloud RoutingLocal MiniCPM classifies each request so sensitive data stays on-deviceLost & Found Desk: Small-Model Item ReturnMiniCPM-V captions items, MiniCPM5 guides claims, staff confirm every handoffCoreAI Model Zoo: MiniCPM on iPhoneCore AI builds of MiniCPM5 and MiniCPM-V 4.6 running natively on iPhone and MacViva Mais: Local WhatsApp CRM CopilotA travel-agency copilot that turns WhatsApp chats into a local-first CRMAranya: Seeing Plants, Telling Their StoriesLoRA-tuned MiniCPM-V 4.6 identifies plants and diagnoses their healthWhen a Catgirl Learns to Think, on Your PhoneWas this model made for catgirls? An on-device deployment testMaking a 1B Model a Good Math AgentTeXada’s on-device MiniCPM practice and lessonsReconciliation Workbench: OCR Tie-OutMiniCPM-V reads financial reports; a deterministic engine catches OCR digit errors