MiniCPM-V 4.7 was formally released and open-sourced by ModelBest on October 13, 2026, in two sizes, 35B-A3B and 1B. The larger one is the first MoE model in the MiniCPM-V line: 35 billion parameters in total, only about 3 billion active per inference, and — the company says — able to run losslessly and smoothly on-device in roughly 3GB of memory. Reports on October 13 from the Global Times, National Business Daily, Jiemian and ITHome agree on the essentials. The real story is not the parameter count but how it tackles the old problem of shuttling experts around on a phone.
The hard part is moving, not computing
An MoE model holds many experts and calls only a few for each answer. The catch is that the idle experts still have to live somewhere: the traditional approach loads an expert's weights into memory only when it is needed, then computes — a stop-and-go rhythm that stutters. The LES (lookahead expert scheduling) technique open-sourced this time takes the opposite approach: the system predicts which experts will be needed next and preloads them while the current computation is still running, overlapping loading with computing. The company calls it the only open-source solution of its kind so far. Also open-sourced is SRFT, a self-replay fine-tuning scheme meant to reduce forgetting when developers adapt the model.
Two speed figures — do not mix them up
The speeds in different reports belong to different devices. ITHome, citing official data, reports the 35B-A3B decoding at over 13 tokens per second on an iPhone 17 Pro Max within about 3GB of memory; the Global Times reports over 15 tokens per second measured on an iPad Pro. The two numbers do not contradict each other, but they should not be swapped. On benchmarks, the company says the 35B-A3B in Instruct mode scores higher overall than Qwen3.6 35B-A3B with shorter inference time, and that the 1B version clearly beats Qwen3.5 0.8B on both score and efficiency. All of these are vendor figures without independent replication so far, and should be read as such.
An adaptation list that ships on day one
According to the company, MiniCPM-V 4.7 supports iOS, Android and HarmonyOS from release day, with chip platforms spanning Rockchip, NVIDIA, Intel and Apple, plus AMD, Arm and Houmo, alongside open-sourced on-device adaptation code and support for mainstream inference frameworks. The MiniCPM series, released by ModelBest with the OpenBMB open-source community, has passed 60 million cumulative downloads. For developers building on-device multimodal apps, the practical meaning is that 35B-class visual understanding now has an open-source option worth seriously trying on phones and tablets, and privacy-sensitive or offline scenarios no longer default to the cloud. Whether it truly runs as losslessly and smoothly as claimed is a verdict to save until the community publishes tests across devices.
