LangBoT has recently rolled out its latest edge-side multimodal large model, MiniCPM-V 4.7, which comes in two variants: 35B-A3B and 1B. The 35B-A3B model employs a Mixture of Experts (MoE) architecture, featuring a staggering 35 billion parameters. However, it only activates roughly 3 billion parameters for each inference, making it the inaugural open-source MoE model to operate seamlessly and without any performance degradation on edge devices. By utilizing proprietary optimization techniques, such as the LES Proactive Expert Scheduling technology, this model achieves a remarkable decoding speed of over 15 tokens per second on an iPad Pro with merely around 3GB of RAM. This represents an approximate fivefold increase in inference speed compared to conventional approaches.

Meanwhile, the 1B lightweight variant further reduces the threshold for edge deployment, accommodating a broader spectrum of devices with constrained configurations. At present, MiniCPM-V 4.7 is compatible with iOS, Android, and HarmonyOS, supports multiple chip platforms, and provides open-source code for edge adaptation.

Statistics reveal that the cumulative global downloads of the MiniCPM series of open-source models, co-launched by LangBoT and the open-source community, have exceeded 60 million. This launch marks a paradigm shift in the technological landscape of edge-side large models, as models boasting hundreds of billions of parameters start to make their way onto mobile terminals.

Nevertheless, transitioning from basic functionality to widespread adoption on consumer-grade smartphones necessitates overcoming hurdles such as power consumption and heat dissipation, multi-platform compatibility, and validation of user experience value. Looking to the future, edge-side large models are expected to take on more intricate multimodal understanding and intelligent agent tasks, thereby diminishing reliance on the cloud.