That day, Mianbi Intelligence launched the new-generation edge-side multimodal large model MiniCPM-V 4.7, offering two sizes in one go: 35B-A3B and 1B. The 35B-A3B is the first MoE (Mixture of Experts) model in the MiniCPM-V family, and relying on the brand-new LES "Lookahead Expert Scheduling" inference acceleration technology, it has become the industry's first open-source MoE that can run losslessly and smoothly on edge devices.
The most eye-catching figure is its real-world performance: on the iPhone 17 Pro Max, it occupies only 3GB of memory while achieving a decoding speed of 13+ token/s. This means there is no need to send data to the cloud — a flagship phone can handle multimodal inference locally, benefiting both privacy and latency.
Cross-platform support, with fine-tuning solutions also open-sourced
In terms of compatibility, 4.7 covers all three major systems — iOS, Android and HarmonyOS — and is compatible with chip platforms including Intel, AMD, Arm, Rockchip and Houmo, encompassing both mobile and local computing scenarios. Mianbi has also released the Self-Replay Fine-Tuning (SRFT) fine-tuning solution, helping developers extend capabilities and suppress forgetting at lower cost, making secondary development easier.
The performance comparison is equally strong: in Instruct mode, 35B-A3B achieves a higher overall score than Qwen3.6 35B-A3B with shorter inference time; compared with the previous generation 4.6, the 1B version's fundamental capabilities continue to climb, with its overall score and inference efficiency clearly surpassing Qwen3.5 0.8B. At a time when edge-side models are competing to get smaller, this generation from Mianbi has pushed the threshold of "large model MoE-ization + running locally on phones" down another notch.