MiniCPM-V 4.7: 35B Mixture-of-Experts Model Reportedly Runs Efficiently on Smartphones

Many accounts welcomed MiniCPM-V 4.7 for its fast local multimodal performance and low ~3 GB memory needs that enable laptop and mobile use, while some replies flagged compression trade-offs that reduce visual detail.

Based on 93 sentiment-bearing replies from 83 accounts across 2 conversations.

OpenBMB’s model card reports 13-plus tokens per second on an iPhone 17 Pro Max using 3GB of memory. It explains that mixture-of-experts models can have more total parameters than they activate for each token. For memory-limited devices, expert weights can stay on SSD or other storage and be loaded when needed, though repeated loading can stall generation.

The project’s Lookahead Expert Scheduling, or LES, predicts which experts the next layer will need and preloads their weights. The model card says prediction only guides loading and scheduling: the actual router still selects which experts run, and missed experts are loaded as usual. Layer-wise quantization reduces weight storage and transfer costs, while DSpark speculative sampling adds another decoding optimization.

The LES guide reports tests on an 11-inch iPad Pro M4 with 16GB of memory. Its unoptimized SSD-loading baseline reached 2.97 tokens per second; scheduling and loading optimizations reached 9.81, adding mixed-precision quantization reached 14.92, and the full setup with DSpark reached 17.39. The guide lists those last two stages as 5.02 and 5.86 times the baseline, respectively. OpenBMB’s launch announcement describes the fivefold speedup as yielding byte-for-byte identical outputs.

Separately, Niels Rogge reported 100 tokens per second locally in an Oct. 17 post describing a llama.cpp web UI demonstration. He did not identify the hardware or model variant for that figure.

Many accounts welcomed MiniCPM-V 4.7 for its fast local multimodal performance and low ~3 GB memory needs that enable laptop and mobile use, while some replies flagged compression trade-offs that reduce visual detail.

Based on 93 sentiment-bearing replies from 83 accounts across 2 conversations.