With the continuous development of artificial intelligence, a new model called MiniCPM-V4.0 has recently attracted widespread attention. This model is the latest version of the MiniCPM-V series, featuring 410 million parameters and built upon SigLIP2-400M and MiniCPM4-3B. Compared to previous versions, MiniCPM-V4.0 shows excellent performance in single-image, multi-image, and video understanding, and has significantly improved efficiency.

image.png

MiniCPM-V4.0 is proud of its strong visual capabilities. In various evaluation benchmarks, the model achieved an average score of 69.0 in the comprehensive assessment by OpenCompass, surpassing models such as GPT-4.1-mini-20250414, MiniCPM-V2.6 (810 million parameters, score 65.2), and Qwen2.5-VL-3B-Instruct (380 million parameters, score 64.5). It also demonstrates good performance in multi-image and video understanding.

Designed for mobile devices, MiniCPM-V4.0 is a major highlight. The model runs smoothly on the iPhone16Pro Max, with a first response delay of less than 2 seconds, a decoding speed of over 17 tokens per second, and no overheating issues. Even under high concurrent requests, it exhibits superior throughput capabilities.

To make it easier for more users to get started, MiniCPM-V4.0 provides multiple usage methods, including tools compatible with various platforms, such as llama.cpp, Ollama, and vLLM. To better serve users, the development team has also open-sourced an iOS application that can run on iPhones and iPads, helping users get started easily. The accompanying Cookbook also provides detailed usage guides and practical examples, further simplifying the operation process.

Project: https://huggingface.co/openbmb/MiniCPM-V-4

Key Points:

🌟 MiniCPM-V4.0 scored 69.0 in the OpenCompass evaluation, surpassing several similar models.  

📱 This model is designed for mobile devices, with fast response times and no overheating issues.  

📚 Open-source iOS application and detailed usage guide, making it easier for users to get started.