Mianbi Intelligence's MiniCPM-V4.6 Open-Source: 1.3B Params Run Smoothly with Only 6G Memory

On May 11, Mianbi Intelligence, together with Tsinghua University and OpenBMB open-source community, released and open-sourced the new-generation edge-side multimodal large model MiniCPM-V4.6. With 1.3B parameters, it runs smoothly on edge devices with only about 6G memory, greatly lowering the threshold for multimodal AI deployment on smart terminals.

In performance, MiniCPM-V4.6 outperforms models of the same size in tasks like general image-text understanding and STEM reasoning, even approaching the level of larger parameter models. For efficiency, its inference throughput is 1.5x that of Qwen3.5-0.8B, while Token consumption is only 2.5% of the latter.

Currently, the model is open-sourced on platforms like GitHub, supports iOS and Android systems, and is compatible with mainstream fine-tuning and inference frameworks. Developers can complete full-process fine-tuning with consumer-grade graphics cards, promoting the popularization of edge-side multimodal applications.