ModelBest (Mianbi Intelligence) released MiniCPM-V 4.7, a new generation on-device multimodal large model, on October 13. The release includes two versions, 35B-A3B and 1B.
The 35B-A3B version uses a Mixture of Experts architecture with 35 billion total parameters and only 3 billion activated per inference. It is the first open-source MoE model capable of running smoothly and without loss on a device, requiring about 3GB of memory and achieving decoding speeds above 15 tokens per second, measured on iPad Pro.
Support was available on launch day for Rockchip, NVIDIA, Intel and Apple chip platforms as well as other mainstream domestic and international ones, across iOS, Android and HarmonyOS, with open-source on-device adaptation code and mainstream inference frameworks provided. The 1B version targets extreme on-device efficiency, bringing flagship-class capability to more devices.
The breakthrough is backed by ModelBest's proprietary LES (Lookahead Expert Scheduling) technology. MoE models contain many internal experts, of which only a few are called for each answer, which places heavy pressure on device memory. The conventional approach loads the required expert on demand, and the loading pause causes stutter. LES predicts in advance which experts will be needed next and pre-loads them while current computation is still running, so loading and computing proceed in parallel. The company describes it as the only open-source solution of its kind currently available in the industry.
ModelBest and the OpenBMB open-source community have published the MiniCPM series of open models with cumulative global downloads exceeding 60 million, covering text, vision and speech modalities.
Source: Global Times / Shanghai Securities News / Daily Economic News
