IT Home On October 13, Mianbi Intelligence announced the launch of a new generation of edge-based multi-modal large model, MiniCPM-V 4.7. It comes in two specifications: 35B-A3B and 1B. The 35B-A3B is the first MoE model in the MiniCPM-V series.
According to official statements, this model is based on the latest LES (Foresight Expert Scheduling) inference acceleration technology. The 35B-A3B version is the first open-source MoE model that supports seamless operation on the edge without loss of performance. It achieves a decoding speed of 13+ tokens/s with only 3GB of memory on the iPhone 17 Pro Max.
IT Home has learned that MiniCPM-V 4.7 supports iOS, Android, and HarmonyOS. It is compatible with chip platforms such as Intel, AMD, Arm, Rockchip, and HiSilicon, covering more mobile devices and local computing scenarios. It implements the open-source Self-Replay Fine-Tuning (SRFT) fine-tuning solution, which helps developers expand model capabilities at low cost and reduce forgetting, making secondary development easier.
In addition, the 35B-A3B version of this model achieves a higher overall score with shorter inference time in the Instruct mode compared to Qwen3.6 35B-A3B; the 1B version continues to improve its basic capabilities over MiniCPM-V 4.6, and is significantly ahead of Qwen3.5 0.8B both in overall score and inference efficiency.
Reference:
Ad notice: The text contains external link references (including but not limited to hyperlinks, QR codes, passwords, etc.) designed to provide more information, save selection time, and the results are for reference only. Articles on IT Home that include external links all contain this notice.