MiniCPM-V 4.7 35B-A3B
A multimodal large model in the MiniCPM-V series released by OpenBMB, and the first open-source on-device MoE model. It has 35B total parameters with about 3B activated per token, supports single-image, multi-image, and video understanding, and can run inference smoothly on on-device hardware such as the iPhone.
Timeline (last 90 days)
ModelBest (OpenBMB) released MiniCPM-V 4.7, the newest member of the MiniCPM-V series, including 35B-A3B, the series' first MoE architecture variant.
MiniCPM-V 4.7 35B-A3B uses a combination of three techniques—Lookahead Expert Scheduling (LES), fine-grained layer-by-layer quantization, and DSpark speculative sampling—to accelerate on-device inference.
MiniCPM-V 4.7 35B-A3B achieves a decoding speed of 13+ tokens/s on the iPhone 17 Pro Max.
In Instruct mode, MiniCPM-V 4.7 35B-A3B achieved a higher overall score with less inference time than Qwen3.6 35B-A3B.
All Facts (4)
MiniCPM-V 4.7 35B-A3B uses a combination of three techniques—Lookahead Expert Scheduling (LES), fine-grained layer-by-layer quantization, and DSpark speculative sampling—to accelerate on-device inference.
50%UnverifiedMiniCPM-V 4.7 35B-A3B achieves a decoding speed of 13+ tokens/s on the iPhone 17 Pro Max.
50%UnverifiedIn Instruct mode, MiniCPM-V 4.7 35B-A3B achieved a higher overall score with less inference time than Qwen3.6 35B-A3B.
50%UnverifiedModelBest (OpenBMB) released MiniCPM-V 4.7, the newest member of the MiniCPM-V series, including 35B-A3B, the series' first MoE architecture variant.
50%