ModelBest open-sources MiniCPM-V 4.7, a 35B-A3B MoE for phones
ModelBest and OpenBMB released MiniCPM-V 4.7 on Oct. 13, a new on-device multimodal model offered in two sizes: 35B-A3B and 1B, with 35B-A3B being the first MoE in the MiniCPM-V family. The company says the 35B-A3B version uses a new LES (look-ahead expert scheduling) inference acceleration technique to decode at 13+ tokens/s on an iPhone 17 Pro Max while occupying only about 3GB of memory, and calls it the first open-source MoE that runs losslessly and smoothly on devices. ModelBest also claims 35B-A3B scores higher overall than Qwen3.6 35B-A3B with shorter inference time in Instruct mode, while the 1B version leads Qwen3.5 0.8B in overall score and inference efficiency. The models support iOS, Android and HarmonyOS and are adapted to Intel, AMD, Arm, Rockchip and Houmo chip platforms, with a Self-Replay Fine-Tuning (SRFT) recipe released alongside.
If the vendor figures hold up, a 35B-class open MoE running smoothly on a phone would lower the barrier for local on-device multimodal apps.