MiniCPM-V 4.5
MiniCPM-V Multimodal Series
MiniCPM-V 4.5 is built on Qwen3-8B and SigLIP2-400M with 8B total parameters, and performs strongly on image / video understanding, OCR and document parsing. It averages 77.0 on OpenCompass across 8 benchmarks — with only 8B parameters it surpasses GPT-4o-latest, Gemini-2.0 Pro and Qwen2.5-VL 72B, making it the most performant MLLM under 30B. It leads GPT-4o-latest and Gemini 2.5 on OCRBench, ranks among the top general MLLMs for PDF document parsing on OmniDocBench, and supports more than 30 languages.
Its three key improvements land in architecture, pre-training and post-training. A unified 3D-Resampler jointly compresses up to 6 consecutive video frames into just 64 tokens (a 96× compression rate), so the model can take in many more frames without extra LLM compute — enabling high-FPS (up to 10FPS) and long video understanding. Pre-training unifies OCR and document knowledge by corrupting text regions and asking the model to reconstruct them, removing the dependence on error-prone document parsers. Post-training uses hybrid reinforcement learning to optimize fast and deep thinking jointly, which is why a single checkpoint can switch between the two modes via enable_thinking.
Looking at this for a new project? Consider MiniCPM-V 4.6 instead — it has a longer context, faster vision tower, and is the actively-recommended release.
Highlights
- 8B parameters (Qwen3-8B + SigLIP2-400M)
- Image, multi-image, and video understanding (no audio)
- One model with optional thinking mode (toggle via
enable_thinking) - HuggingFace:
openbmb/MiniCPM-V-4_5· ModelScope:OpenBMB/MiniCPM-V-4_5
Related Resources
| Topic | Link |
|---|---|
| Inference | MiniCPM-V 4.5 inference guide |
| Deployment | MiniCPM-V 4.5 deployment guide |
| Quantization | MiniCPM-V 4.5 quantization guide |
| Code repository | OpenBMB/MiniCPM-V |
| Technical report | MiniCPM-V 4.5 Technical Report |