OpenBMB Releases MiniCPM-V-4.7-35B-A3B Open-Source MoE Vision-Language Model
OpenBMB has released MiniCPM-V-4.7-35B-A3B, the first open-source Mixture-of-Experts vision-language model optimized for lossless on-device inference. It features innovative techniques like Lookahead Expert Scheduling and DSpark speculative sampling to achieve smooth performance on mobile and edge hardware. This release advances the performance-efficiency Pareto frontier for multimodal models, enabling powerful video and image understanding on resource-constrained devices like smartphones. It significantly lowers the barrier for running sophisticated multi-modal AI locally across diverse consumer and enterprise platforms. The 35B-A3B variant achieves a decoding speed exceeding 13 tokens per second on an iPhone 17 Pro Max while supporting mixed 4x/16x visual token compression. It offers extensive framework and hardware support, including day-0 compatibility for chips from Rockchip, Intel, AMD, and Arm, alongside GGUF, AWQ, and FP8 quantization variants.
## BACKGROUND
Mixture-of-Experts (MoE) is a machine learning architecture where only a subset of neural network parameters (experts) is activated for any given input, reducing compute costs. Vision-language models (MLLMs) integrate visual encoders with language models to enable tasks like image captioning, visual question answering, and video understanding.