MiniCPM-V 4.7 brings a 35B multimodal MoE model to phones with about 3 GB of memory
ModelBest and the OpenBMB community have released MiniCPM-V 4.7, an open-source on-device multimodal model series in 35B-A3B and 1B versions. According to the companies, it brings image, document and video understanding to edge devices with about 3GB of memory.
- ModelBest and OpenBMB released MiniCPM-V 4.7 in 35B-A3B and 1B versions, targeting devices with about 3GB of memory.
- The 35B-A3B model activates about 3 billion parameters per token, with most experts kept on SSD and loaded as needed.
- Lookahead Expert Scheduling prepares more than 80% of required experts in advance, and outputs are identical with it on or off.
- The model reaches more than 17 tokens per second on an iPad Pro M4 and more than 13 on the iPhone 17 Pro Max.
- Weights, inference code and fine-tuning code are publicly available, with Day-0 support from multiple chipmakers.
A mixture-of-experts model designed for phones
ModelBest describes MiniCPM-V 4.7 35B-A3B as the first open-source MoE multimodal model built to run smoothly on-device without compromising model quality. The model activates about 3 billion parameters per token, while most experts stay on SSD storage and are loaded as needed. This is how the release brings large-model multimodal capability to iOS and Android.
The model covers general visual understanding, STEM problem solving, OCR, document and chart interpretation, and long-video understanding. According to ModelBest, it pushes forward the intelligence-cost Pareto frontier for on-device multimodal AI by delivering strong overall performance with short inference time.
The 1B version builds on MiniCPM-V 4.6 and targets lightweight edge devices. It supports general vision, OCR, chart reasoning and multi-image understanding, and can run smoothly on entry-level devices, which the companies say brings multimodal AI to more everyday hardware.
Lookahead scheduling and speed figures
A key part of the speed gains is ModelBest's Lookahead Expert Scheduling (LES). It predicts which experts the next layer will need and loads them while the current layer is computing. In most scenarios, it prepares more than 80% of required experts in advance without additional training. The model's original router still makes the final expert selection, so outputs are identical whether LES is enabled or disabled.
LES works together with layer-wise quantization and contiguous expert storage. Combined, these optimizations increased average generation speed approximately 5x on an iPad Pro M4 with 16 GB of memory. The model reaches more than 17 tokens per second on that device and more than 13 tokens per second on the iPhone 17 Pro Max.
The model also supports both 4× and 16× visual token compression, so developers can prioritize detail or speed. At 16× compression, it scores 90.9 on MMBench and 90.4 on OCRBench, according to the release.
Fine-tuning method and hardware support
The release introduces Self-Replay Fine-Tuning, which mixes the model's own high-quality responses into training data. The aim is to help the model learn new tasks while retaining its general abilities, at a cost close to standard supervised fine-tuning.
MiniCPM-V 4.7 has Day-0 support from AMD, Apple, Arm, Houmo AI, Intel, Rockchip and NVIDIA. Through the FlagOS software stack, it also has Day-0 support for AI and edge chips from CIX Technology, Huawei, Hygon, Iluvatar CoreX, Kunlunxin, MetaX, Moore Threads, Sunrise, T-Head and TsingMicro.
What it means for developers
Model weights, inference code and fine-tuning code are publicly available for developers to download, evaluate and deploy. For teams building mobile or edge applications, the release offers a large multimodal model that, according to its makers, fits within about 3GB of memory, along with a smaller 1B option for entry-level hardware. The speed and benchmark figures come from ModelBest's own announcement.
