IT Home Releases MiniCPM-V 4.7 AI Model for On-Device Use

IT Home announced on October 13 the release of its new generation on-device multimodal large language model, MiniCPM-V 4.7. The model is available in two versions, 35B-A3B and 1B, with the 35B-A3B being the first MoE (Mixture of Experts) model in the series.

According to the company, the model utilizes the latest LES (Look-ahead Expert Scheduling) inference acceleration technology. The 35B-A3B version is highlighted as the first open-source MoE model capable of running losslessly and smoothly on-device. IT Home reports a decoding speed exceeding 13 tokens per second using only 3GB of memory, demonstrated on devices like the iPhone 17 Pro Max, enabling complex AI tasks directly on mobile hardware.

The MiniCPM-V 4.7 model is compatible with iOS, Android, and HarmonyOS, and supports various chip platforms including Intel, AMD, Arm, and Rockchip. This broad compatibility aims to cover a wider range of mobile devices and local computing scenarios.

Additionally, IT Home has open-sourced the Self-Replay Fine-Tuning (SRFT) method. This aims to help developers expand model capabilities at a lower cost and reduce forgetting during fine-tuning, facilitating further development and application of the model. The company claims the 35B-A3B version outperforms competitors like Qwen3.6 35B-A3B in Instruct mode, while the 1B version shows significant improvements over Qwen3.5 0.8B.