On the same day, Mingshi Intelligence launched its new generation of edge-side multimodal large model MiniCPM-V 4.7, offering two specifications: 35B-A3B and 1B. Among them, 35B-A3B is the first MoE (Mixture of Experts) model in the MiniCPM-V family, and it relies on the newly developed "forward-looking expert scheduling" inference acceleration technology, becoming the first open-source MoE that can run without loss and smoothly on the edge.
The most eye-catching number is the performance: on an iPhone 17 Pro Max, it only takes up 3GB of memory to achieve a decoding speed of 13+ token/s. This means that data does not need to be sent to the cloud, and high-end smartphones can handle multimodal reasoning locally, benefiting from both privacy and latency.
Supporting multiple platforms, with fine-tuning solutions also open-sourced
In terms of compatibility, version 4.7 covers three major systems: iOS, Android, and HarmonyOS, and is compatible with chip platforms such as Intel, AMD, Arm, Rockchip, and Huanmo, covering both mobile and local computing scenarios. Mingshi also released the Self-Replay Fine-Tuning (SRFT) fine-tuning solution, helping developers expand capabilities at a lower cost and suppress forgetting, making secondary development more convenient.
The performance comparison is also strong: in Instruct mode, the 35B-A3B outperforms Qwen3.6 35B-A3B with shorter inference time and higher overall score; the 1B version continues to improve upon the previous version 4.6, with significant advantages in overall score and inference efficiency over Qwen3.5 0.8B. In the current context where edge-side models are getting smaller and more competitive, Mingshi's latest version has further lowered the threshold for "MoE for large models + running on smartphones locally."
