MiniCPM-V 4.7 35B-A3B Hits 30+ Token/s on Phones, ModelBest Says

On October 13, 2026, OpenCPM released and open-sourced the MiniCPM-V 4.7 series in two variants, 35B-A3B and 1B. On November 5, the company held an on-device multimodal LLM technical exchange event, where Yao Yuan, OpenCPM's co-chief scientist and an assistant professor at the School of Artificial Intelligence at Tsinghua University, said the team had pushed the on-phone generation speed of MiniCPM-V 4.7 35B-A3B above 30 Token/s in its latest inference optimization round. The phone-side deployment reportedly uses only about 3GB of memory, with almost negligible accuracy loss. The model is a Mixture-of-Experts (MoE) design with 35 billion total parameters, of which roughly 3 billion are activated per inference, putting it close to smooth real-time interaction on a smartphone. OpenCPM co-founder and chief scientist Liu Zhiyuan said at the same event that shipping a large model into a terminal is not enough; the key is getting users to actually use it. He argued that a generation speed of at least 30 Token/s is needed to deliver a sufficiently smooth experience. Afterwards, Yao and core members of OpenCPM's multimodal lab, Yu Tianyu and Cai Tianchi, presented the team's work from three angles: trends in on-device intelligence, model capability density, and low-level inference optimization. Beyond what was covered at the event, no information was given on the broader release of this optimization or when it might reach general devices.