ModelBest Open-Sources MiniCPM-V 4.7: First On-Device MoE Achieves 13+ Tokens/s in 3GB RAM via LES Scheduling

ModelBest and OpenBMB have officially open-sourced MiniCPM-V 4.7, marking a significant breakthrough in on-device AI deployment. The flagship 35B-A3B version introduces a Mixture-of-Experts (MoE) architecture for the first time in this series, activating only 3B parameters to compress memory usage to approximately 3GB. This innovation enables high-performance multimodal models to run fully offline on mainstream mobile devices like the iPhone 17 Pro Max, achieving optimized decoding speeds exceeding 13 tokens per second.

Cross-verified reports highlight that the model's core advantage lies in its novel LES (Lookahead Expert Scheduling) technology, which addresses inference latency issues of MoE architectures on resource-constrained devices. MiniCPM-V 4.7 demonstrates superior performance in benchmarks such as MMBench and MMMU, capable of complex chart analysis and precise visual question answering. Additionally, the company released the Self-Replay Fine-Tuning (SRFT) scheme to help developers expand capabilities cost-effectively while mitigating catastrophic forgetting. The model supports extensive platform compatibility, including iOS, Android, HarmonyOS, and chips from Intel, AMD, Arm, and Rockchip, integrating seamlessly with frameworks like vLLM and llama.cpp.

This release validates the feasibility of high-performance small-parameter models under the 'Density Law' and reshapes the landscape of on-device multimodal AI. By providing a full-stack open-source solution from underlying architecture to fine-tuning tools, ModelBest accelerates developer ecosystem construction, driving the widespread adoption of intelligent, low-power AI applications across smartphones and IoT devices.