Mingshi Intelligence launched MiniCPM-V 4.7, featuring 35B-A3B and 1B specs. The 35B-A3B is the first open-source MoE model running losslessly on edge devices using forward-looking expert scheduling. It operates on an iPhone 17 Pro Max with just 3GB memory at 13+ token/s decoding speed, enabling local multimodal reasoning for enhanced privacy and latency. Compatible with iOS, Android, and HarmonyOS across Intel, AMD, Arm, Rockchip, and Huanmo chips, it also offers open-source Self-Replay Fine-Tuning. The 35B-A3B outperforms Qwen3.6 35B-A3B in speed and scores, while the 1B version surpasses Qwen3.5 0.8B, lowering thresholds for large models on smartphones.