Face Intelligence Unveils MiniCPM-V 4.7 Edge Model with 13+ tok/s On-Device Speed

On October 13, Face Intelligence announced MiniCPM-V 4.7, its next-generation on-device multimodal large language model, available in 35B-A3B and 1B variants. The 35B-A3B is the series' first mixture-of-experts (MoE) model, built on the company's latest LES (Lookahead Expert Scheduling) inference acceleration technology. It achieves a decoding speed of over 13 tokens per second on an iPhone 17 Pro Max while using only 3GB of memory, and is described as the first open-source MoE model that runs losslessly on edge devices.

MiniCPM-V 4.7 supports iOS, Android, and HarmonyOS, and is compatible with chips from Intel, AMD, Arm, Rockchip, and Houmo, covering a wider range of mobile and local computing scenarios. The 1B variant continues to improve basic capabilities over MiniCPM-V 4.6. The company also open-sourced the Self-Replay Fine-Tuning (SRFT) scheme to help developers expand model capabilities at lower cost and reduce forgetting.

According to official data, the 35B-A3B variant achieved a higher overall score than Qwen3.6 35B-A3B in Instruct mode with shorter inference time, while the 1B variant significantly outperformed Qwen3.5 0.8B in both overall score and inference efficiency.