MiniCPM-V 4.7 puts a 35B vision MoE on phones, OpenBMB says
OpenBMB released a 35B-A3B multimodal model and a 1B variant; its headline iPhone speed and memory figures are the developer's own measurements.
A large vision model aimed at local devices
OpenBMB’s October 12 release of MiniCPM-V 4.7 includes a 35B-A3B mixture-of-experts model and a smaller 1B variant. The English repository and model card describe image, multi-image and video understanding, downloadable weights, and guides for local deployment. The larger model is the first mixture-of-experts entry in the MiniCPM-V series. Its release is a notable community option for developers who want multimodal work on their own hardware rather than a hosted-only API.
The phone claim has a source, not an independent test
OpenBMB says Lookahead Expert Scheduling preloads likely expert weights to reduce stalls while retaining token-level routing. The project reports more than 13 tokens per second on an iPhone 17 Pro Max with 3 GB of memory used for its on-device setup. That speed, memory figure, and its comparisons with other models are OpenBMB’s own results; they do not establish performance across every phone or task. The 35B-A3B model was observed on Hugging Face’s trending list on October 13, after the October 12 release. That current visibility does not prove a new release on October 13 or a measured surge in adoption.