The Pulse
OpenBMB Says Its 35B Vision Model Runs on iPhone at 13 Tokens a Second
OpenBMB released MiniCPM-V 4.7 on October 12, including a 35B-A3B mixture-of-experts vision model and a smaller 1B version. The company reports speeds above 13 tokens per second on an iPhone 17 Pro Max, while its phone-app guide calls for 1
AI.info Team · Written with AI ·
More than 13 tokens per second on an iPhone is OpenBMB’s headline claim for MiniCPM-V 4.7, a vision-language model with 35 billion total parameters. The company says its 35B-A3B mixture-of-experts version reached that decoding speed on an iPhone 17 Pro Max while using 3 GB of memory. OpenBMB released the model on October 12, alongside a separate 1B version.
How OpenBMB gets a 35B model onto a phone
MiniCPM-V 4.7 handles images, multiple images, documents and video, according to its release documentation. OpenBMB describes the 35B-A3B version as the first MoE model in the MiniCPM-V family; the “A3B” designation refers to 3 billion active parameters, rather than all 35 billion being used for each token. The model also supports mixed 4× and 16× visual-token compression, which lets developers trade image detail for lower computation.
The speed claim rests on an inference setup designed around the constraints of phone storage and memory. OpenBMB says the model is “achieving a decoding speed of 13+ tokens/s with just 3 GB of memory on iPhone 17 Pro Max.” The company credits Lookahead Expert Scheduling, which predicts and preloads the experts the next layer is likely to need, along with weight quantization and an optimized storage layout that shrink the footprint and speed up loading.
The 3 GB figure is not the full phone requirement
OpenBMB’s companion app repository adds an important qualification: the phone deployment streams expert weights from an SSD instead of keeping them all in memory. Its guide lists 12 GB of device RAM as the recommendation for the 35B-A3B version and says the setup has been validated on a 12 GB iPhone 17 Pro Max. The guide estimates a 16.6 GB download and about 17 GB of free storage, including the main model file, vision projector and 40 expert shards.
That distinction matters: the 3 GB figure in the model announcement is not a claim that the full model download fits in 3 GB of phone storage, or that any 3 GB phone can run it. The phone app uses a split, on-device package and streams expert weights from storage; OpenBMB says phones with less RAM fall back to a smaller expert cache and run considerably more slowly. Its guide also says that the special split package is made for the demo, rather than a general-purpose model release.
OpenBMB pairs the MoE with a 1B model
The smaller MiniCPM-V 4.7 variant is a 1B dense model, not an MoE. OpenBMB says it outperforms its earlier MiniCPM-V 4.6 and Qwen3.5 0.8B on the company’s reported performance and efficiency comparisons, while the 35B-A3B model scores higher and runs in less inference time than Qwen3.6 35B-A3B in Instruct mode. The release documentation does not provide numerical benchmark results in its text, so those comparisons remain the developers’ claims.
The release supports deployment on iOS, Android and HarmonyOS, and lists chip support spanning Arm, Intel, AMD, Rockchip and Houmo. OpenBMB says the code and model weights are available under the Apache-2.0 license. Its companion app repository provides sample phone apps and the specialist storage arrangement needed to test the 35B model locally.