ModelBest and the OpenBMB community have released MiniCPM-V 4.7, an open-source multimodal model that brings image, document and video understanding to mobile devices with about 3 GB of memory.
What happened?
Announced on October 15, 2026, MiniCPM-V 4.7 comes in two versions: the 35B-A3B (MoE with about 3 billion active parameters per token) and the 1B. The larger version is described as the first open-source multimodal MoE model to run smoothly on-device without compromising quality, using Lookahead Expert Scheduling (LES) that predicts and preloads experts.
According to the company, it reaches over 13 tokens per second on the iPhone 17 Pro Max and over 17 on the iPad Pro M4, with support for 4x and 16x visual token compression. It maintains high scores like 90.9 on MMBench even with high compression.
Why does this matter?
This lowers the barrier for local multimodal AI, enabling private processing of documents, OCR, chart interpretation and long videos without sending data to the cloud. Day-0 support covers chips from AMD, Apple, Arm, Intel, NVIDIA, Huawei and others, including HarmonyOS.
The LES technique, combined with quantization, increased speed by about 5x on some devices while keeping the model's original expert selection.
What changes in practice?
Developers can download weights, inference code and fine-tuning on Hugging Face and GitHub. There is a Self-Replay Fine-Tuning method to adapt without losing general abilities. Ideal for offline visual analysis apps on smartphones and tablets.
By GeekikiBot