OpenBMB Releases MiniCPM-V 4.7: First On-Device MoE Multimodal Model Achieving 13+ Tokens/s on 3GB RAM

October 13 — ModelBest announced the launch of its new generation on-device multimodal large model, MiniCPM-V 4.7, available in two specifications: 35B-A3B and 1B. Notably, the 35B-A3B variant is the first MoE (Mixture of Experts) model in the MiniCPM-V series.

According to official information, this model leverages the latest LES (Lookahead Expert Scheduling) inference acceleration technology. The 35B-A3B version is the first open-source MoE model capable of running smoothly and losslessly on edge devices, achieving a decoding speed of 13+ tokens/s with only 3GB of memory.

IT Home reports that MiniCPM-V 4.7 supports iOS, Android, and HarmonyOS, and is adapted for chip platforms including Intel, AMD, Arm, Rockchip, and Houmo, covering a wider range of mobile devices and local computing scenarios. Additionally, the Self-Replay Fine-Tuning (SRFT) method has been open-sourced simultaneously, helping developers expand model capabilities at low cost while mitigating catastrophic forgetting, thereby facilitating secondary development.

Furthermore, in Instruct mode, the 35B-A3B version achieves higher overall scores than Qwen3.6 35B-A3B with shorter inference times. The 1B version continues to improve upon MiniCPM-V 4.6's foundational capabilities, significantly outperforming Qwen3.5 0.8B in both overall score and inference efficiency.

Reference:

• MiniCPM-V/README_zh.md at main · OpenBMB/MiniCPM-V · GitHub