Skip to main content

MiniCPM-V 4.7 Runs on Your Phone—No Cloud Needed

MingTang Intelligent's MiniCPM-V 4.7 Brings Big AI to Small Devices

Imagine running a powerful multimodal AI model right on your smartphone—no cloud connection, no data leaving your device. That's exactly what MingTang Intelligent has just made possible with MiniCPM-V 4.7, a new generation of edge-side models announced today.

The lineup includes two versions: 35B-A3B and 1B. The 35B-A3B is the first MoE (Mixture of Experts) model in the MiniCPM-V family. Thanks to a clever new technique called "forward-looking expert scheduling," it becomes the first open-source MoE that runs losslessly and smoothly on edge devices.

Image

Performance That Turns Heads

Here's the number that really grabs attention: on an iPhone 17 Pro Max, the model takes up just 3GB of memory and hits a decoding speed of 13+ tokens per second. That means your phone can handle multimodal reasoning locally—keeping your data private and cutting out latency. No more waiting for the cloud to respond.

Broad Compatibility and Easy Fine-Tuning

Version 4.7 works across iOS, Android, and HarmonyOS, and supports chips from Intel, AMD, Arm, Rockchip, and Huanmo. So whether you're on a phone or a local computer, you're covered.

MingTang also released the Self-Replay Fine-Tuning (SRFT) solution. This helps developers expand model capabilities at a lower cost while suppressing forgetting—making secondary development much more convenient.

How It Stacks Up

The performance comparisons are impressive. In Instruct mode, the 35B-A3B beats Qwen3.6 35B-A3B with shorter inference time and a higher overall score. The 1B version improves on the previous 4.6, showing clear advantages over Qwen3.5 0.8B in both overall score and inference efficiency.

As edge-side models get smaller and the competition heats up, MingTang's latest release further lowers the barrier for running large MoE models locally on smartphones.

Key Points

  • MiniCPM-V 4.7 comes in two sizes: 35B-A3B and 1B.
  • 35B-A3B is the first open-source MoE that runs losslessly on edge devices, thanks to "forward-looking expert scheduling."
  • On an iPhone 17 Pro Max, it uses only 3GB memory and achieves 13+ tokens/s.
  • Supports iOS, Android, HarmonyOS and major chip platforms.
  • SRFT fine-tuning solution is open-sourced for easier development.
  • Outperforms comparable models from Qwen in speed and overall score.