OpenBMB Releases MiniCPM-V 4.7 Enabling 35B MoE Model on Mobile
OpenBMB has introduced MiniCPM-V 4.7, featuring a 35B-A3B Mixture-of-Experts (MoE) vision-language model that runs locally on mobile devices using approximately 3GB of memory. The release also includes a smaller 1B variant for resource-constrained devices and introduces Self-Replay Fine-Tuning (SRFT) for task customization. This breakthrough demonstrates that large multimodal models can operate locally on edge devices without relying on cloud infrastructure, enhancing user privacy and reducing latency. It marks a significant step forward in making advanced AI assistants practical for everyday consumer hardware. The model achieves high efficiency using Lookahead Expert Scheduling (LES), which predicts and preloads expert weights to eliminate bottlenecks during on-device execution. Additionally, Self-Replay Fine-Tuning (SRFT) allows developers to customize the model for specific tasks while preserving its general capabilities.
## BACKGROUND
Mixture-of-Experts (MoE) is a machine learning architecture where only a subset of neural network parameters, known as experts, are activated for any given input, enabling high model capacity with lower computational costs. Vision-language models combine computer vision and natural language processing to interpret and reason about images alongside text.