MiniCPM-V 4.7 35B-A3B: the series' first MoE model with on-device inference

On October 7, 2026, the model openbmb/MiniCPM-V-4.7-35B-A3B was published on HuggingFace with the image-text-to-text task. This is the first MoE architecture variant in the MiniCPM-V line: 35.2 billion parameters, of which not all are active at once. The author claims a decoding speed of 13+ tokens per second on an iPhone 17 Pro Max with 3 GB of memory thanks to the Lookahead Expert Scheduling technique and layer-wise quantization.

What it means

For running on your own hardware, the model is available in safetensors format via transformers, license — apache-2.0. The full repository weighs 65.6 GB, the largest file — 4.7 GB. The platform already has 10 quantized builds: GGUF, FP8, GPTQ, AWQ and MLX. The lightest working build is IQ4_XS by AtomicChat at 16.4 GB; we don't count builds worse than 4 bit. In parallel, MiniCPM-V-4.7-1B was released in the same days, but its fact sheet is not in the news note. You can follow the appearance of new models and quantized builds in the CPU3D news feed.