Model Best Releases MiniCPM-V 4.7 Edge Multimodal Model
Model Best released MiniCPM-V 4.7, an edge-optimized multimodal model available in 35B-A3B and 1B sizes, utilizing the LES reasoning acceleration technology. The 35B-A3B version is the first open-source MoE model capable of running smoothly on mobile devices, achieving over 13 tokens per second on an iPhone 17 Pro Max using only 3GB of memory. This release represents a significant breakthrough for on-device AI by proving that large-scale Mixture of Experts models can run efficiently with minimal memory footprints on everyday hardware. It broadens the deployment scope of multimodal artificial intelligence across diverse mobile operating systems and hardware chips. The model natively supports iOS, Android, and HarmonyOS, alongside various chip platforms like Intel, AMD, and ARM. Additionally, it open-sources the Self-Replay Fine-Tuning (SRFT) scheme to help developers easily expand model capabilities and minimize catastrophic forgetting during secondary development.
## BACKGROUND
Edge AI refers to artificial intelligence models running directly on local consumer devices like smartphones and laptops, offering low latency, enhanced privacy, and offline functionality. Mixture of Experts (MoE) is a model architecture where only a subset of network parameters or 'experts' are activated for any given input, drastically reducing computational overhead.