ModelBest Releases MiniCPM-V 4.7 for Phones and Tablets

ModelBest and OpenBMB released MiniCPM-V 4.7, the companies announced in a press release. The open source multimodal model series includes 35B-A3B and 1B versions for processing images, documents and video on mobile devices.

The 35B-A3B mixture of experts model activates about 3 billion parameters for each token. Most experts remain in SSD storage and load when needed, allowing the model to operate on iOS and Android devices with about 3 GB of memory. It supports visual understanding, OCR, charts, documents, science and mathematics problems, and long videos.

A scheduling system predicts and loads experts before they are required. Combined with quantization and contiguous expert storage, it increased average generation speed about fivefold on an iPad Pro M4 with 16 GB of memory, according to ModelBest. The model generated more than 17 tokens per second on that device and more than 13 tokens per second on an iPhone 17 Pro Max.

The smaller 1B version supports vision, OCR, chart reasoning and analysis across multiple images. Model weights, inference code and fine tuning code are publicly available, with launch support for hardware from AMD, Apple, Arm, Intel, NVIDIA and other chip suppliers.