OpenBMB · 2026-10-12 · notable
MiniCPM-V 4.7 — OpenBMB's first MoE vision model runs on an iPhone
MiniCPM-V 4.7 is an Apache-2.0 vision-language family from OpenBMB: a 35B-A3B mixture-of-experts model that decodes at 13+ tokens/s on an iPhone 17 Pro Max in about 3 GB, plus a 1B model.
MiniCPM-V 4.7 brings a 35B mixture-of-experts vision model down to a phone's memory budget.
Key specs
| Omni doc bench v1.6 | 94.24 |
|---|---|
| I phone 17 pro max | 13+ tok/s |
What is it?
MiniCPM-V 4.7 adds the first mixture-of-experts model to OpenBMB's MiniCPM-V series, a 35B-A3B that reads single images, several images and video, alongside a 1B model. Weights and code are released under Apache-2.0 on Hugging Face.
How does it work?
A SigLIP2 vision encoder feeds a Qwen3.5/Qwen3.6-family language backbone, and each image can be squeezed into 4x or 16x fewer visual tokens. Lookahead Expert Scheduling preloads the experts the next layer will likely need, which OpenBMB says gives a 5x speedup on real devices.
Why does it matter?
Local vision assistants usually rely on small dense models. In OpenBMB's own instruct-mode table the 35B-A3B leads Qwen3.6-35B-A3B, Qwen3.8-27B and Gemma-4-31B on MMBench, MMMU and OmniDocBench v1.6, while running in about 3 GB on a phone.
Who is it for?
on-device and edge AI developers
Try it
https://huggingface.co/spaces/openbmb/MiniCPM-V-4.7-Demo