ModelBest and OpenBMB Unveil MiniCPM-V 4.7 for Enhanced Mobile AI Performance

Revolutionizing Mobile AI: The Launch of MiniCPM-V 4.7

On October 15, 2026, ModelBest, alongside the OpenBMB community, officially launched MiniCPM-V 4.7, a significant advancement in the realm of open-source on-device AI models. With its remarkable 35 billion parameters dedicated to various tasks, including image and video understanding, this new addition is designed to perform efficiently on mobile devices without compromising quality.

What Makes MiniCPM-V 4.7 Unique?

MiniCPM-V 4.7 is a milestone for ModelBest as it introduces the first-ever open-source Multimodal Mixture of Experts (MoE) model that runs seamlessly on mobile and edge devices equipped with approximately 3GB of memory. The innovative design ensures that over 3 billion parameters per token are activated selectively, enhancing processing proficiency. This model paves the way for enhanced visual understanding in general contexts, handling tasks ranging from STEM problem-solving to real-time document interpretation and long-video comprehension.

Addressing Mobile Challenges

The greatest challenge for mobile AI has been balancing efficiency and performance. ModelBest’s breakthrough facilitates this by optimizing how the model accesses its parameters. It utilizes a technique called Lookahead Expert Scheduling (LES), which intelligently predicts which parameters will be required next while the current layer is executing, thus preparing them ahead of time. This handy feature means that in typical scenarios, more than 80% of the needed parameters are loaded even before they are required.

Impressive Speed for Heavy Tasks

What’s truly impressive about the MiniCPM-V 4.7 is its remarkable speed. Benchmarked on an iPad Pro M4 with 16GB of memory, the model achieves an average generation speed that is approximately five times faster than previous versions. This translates to above 17 tokens per second on the iPad and over 13 on the iPhone 17 Pro Max. Users can therefore expect a rapid response and high-level performance, making advanced AI more accessible and practical on everyday devices.

Versatility for Different Devices

In response to the increasing need for scalable AI solutions, the release includes a lightweight 1B version aimed at entry-level devices. This version retains core functionalities like OCR and multi-image understanding while ensuring efficient performance on less powerful hardware. This layered approach allows developers to cater to a wider audience, empowering consistent access to advanced AI capabilities.

Benchmark Testing and Optimization

The MiniCPM-V 4.7 performs remarkably well on benchmark tests, scoring above 90 points across various metrics like MMBench and OCRBench, even under extreme visual token compression settings. It facilitates flexibility for developers, allowing them to choose between higher detail and increased operational speed based on application needs.

Continuous Learning through Self-Replay

To maintain its competitive edge and versatility, MiniCPM-V 4.7 introduces a Self-Replay Fine-Tuning method. This innovative technique incorporates high-quality responses of the model into its training dataset, ensuring that it learns new tasks effectively while preserving its foundational abilities. Notably, this approach introduces minimal additional training costs, aligning closely with conventional supervised training costs.

Industry Collaboration and Future Outlook

MiniCPM-V 4.7 boasts substantial support from tech giants such as AMD, Apple, Arm, and NVIDIA, among others. This collaboration ensures that robust support is in place for various AI and edge computing technologies, paving the way for future innovation in terms of AI applications on mobile devices. With publicly available model weights and inference codes, developers are invited to explore, evaluate, and deploy solutions powered by this cutting-edge technology.

In conclusion, the launch of MiniCPM-V 4.7 marks a new era in mobile AI technology, where high-performance multimodal capabilities are now within reach of everyday devices. As we move forward with this innovation, one can only envision the limitless potential of AI effectively integrated into our daily lives.