OpenBMB's MiniCPM-V 4.7 Runs a 35B Vision Model on Your Phone
OpenBMB's MiniCPM-V 4.7 packs 35B parameters into a mixture-of-experts vision model that runs on an iPhone at 13+ tokens per second.
- OpenBMB released MiniCPM-V 4.7, the first MoE model in the MiniCPM-V series, with 35B total and 3B active parameters.
- Runs at 13+ tokens/sec on iPhone 17 Pro Max using only 3 GB of memory, claimed as the first smooth on-device open-source MoE.
- Lookahead Expert Scheduling (LES) predicts and prefetches experts across layers to overlap loading with compute.
- Beats Qwen3.6 35B-A3B on overall score with lower inference time, per OpenBMB's benchmarks.
- Apache 2.0 license, with FP8, INT4, AWQ, GPTQ and GGUF quantizations plus vLLM, SGLang, llama.cpp support.
- Day-0 chip support for Rockchip, Houmo, Intel, AMD, Arm; iOS, Android, HarmonyOS deployments via edge apps repo.
MiniCPM-V 4.7 targets phones with a 35B MoE model
OpenBMB has released MiniCPM-V 4.7 weights, an Apache 2.0 vision-language model designed to run across servers, desktops, and mobile devices. OpenBMB reports that the 35B-A3B variant generates more than 13 tokens per second using about 3 GB of memory on an iPhone 17 Pro Max. The result matters because large mixture-of-experts models usually depend on server-class memory capacity and bandwidth.
Thirty-five billion parameters, three billion active
The 35B-A3B checkpoint is the MiniCPM-V series’ first mixture-of-experts model. It contains 35 billion total parameters and activates roughly 3 billion for each token. The active subset limits computation per token, while the full parameter count determines checkpoint size and the volume of weights that may need to move from storage.
Raw BF16 storage for 35 billion parameters requires roughly 70 GB before metadata and runtime overhead. Nominal four-bit storage reduces that figure to about 17.5 GB. OpenBMB’s reported 3 GB result therefore describes runtime memory within a design that keeps quantized weights on storage and loads experts as needed.
- Inputs: text, single images, multiple images, and video
- Architecture: 35 billion total parameters with about 3 billion active per token
- Vision encoder: SigLIP2 with early visual-token compression
- Language backbone: components from the Qwen3.5 and Qwen3.6 families
- License: Apache 2.0
How expert paging stays ahead
Mixture-of-experts routing creates an awkward mobile workload. Each token can select different experts at different layers, yet a phone may lack enough RAM to retain every expert. Loading selected weights from flash storage during each layer can leave the processor waiting, while retaining a fixed expert subset restricts token-level routing.
OpenBMB combines three techniques to reduce those stalls:
- Lookahead Expert Scheduling:
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
