OpenBMB is a research group behind the MiniCPM model series. Its recent coverage centers on on-device inference: it released MiniCPM-V 4.7-35B-A3B, the first MoE model in the MiniCPM-V line, running at over 13 tokens/s on an iPhone 17 Pro Max; demoed the 35B vision-language model locally with about 3GB memory alongside a 1B variant; and detailed Lookahead Expert Scheduling, which runs a 35B-A3B MoE model at 20.51 tok/s on an iPhone 18 Pro Max, 6.07x faster than baseline.
OpenBMB runs 35B MoE model locally on iPhoneOpenBMB detailed an engineering approach that runs a 35B-A3B MoE model at 20.51 tok/s on an iPhone 18 Pro Max, 6.07x faster than baseline. Its Lookahead Expert Scheduling preloads predicted expert weights to hide storage bottlenecks without skipping experts or changing routing, using about 3 GB peak memory; the code is open source.
MiniCPM-V 4.7 runs 35B multimodal model on phoneOpenBMB demoed MiniCPM-V 4.7, a 35B-A3B MoE vision-language model running locally on phones with roughly 3GB memory, using Lookahead Expert Scheduling to cut bottlenecks, plus a smaller 1B variant.
OpenBMB releases MiniCPM-V 4.7-35B-A3B multimodal modelOpenBMB released MiniCPM-V 4.7-35B-A3B, the first MoE model in the MiniCPM-V series, enabling smooth on-device inference at over 13 tokens/s on iPhone 17 Pro Max and advancing the performance-efficiency Pareto frontier beyond Qwen3.6 35B-A3B.
The A.I. Daily