OpenBMB releases MiniCPM-V 4.7, series' first MoE model

OpenBMB released MiniCPM-V 4.7 on Hugging Face this week, with a 35B-A3B mixture-of-experts variant the team describes as "the first MoE model in the MiniCPM-V series." The headline efficiency claim, on the MiniCPM-V 4.7 collection, is 13+ tokens/s with just 3 GB of memory on iPhone 17 Pro Max — unusual territory for a model whose total parameter count is 35 billion.

The MoE model ships alongside a 1B companion, both built on a Qwen3.5/Qwen3.6 language backbone with a SigLIP2 vision module. On the 35B card, OpenBMB says the Instruct variant "achieves a higher overall score in less inference time than Qwen3.6 35B-A3B," and claims the 1B variant "substantially outperforms Qwen3.5 0.8B in both overall score and inference efficiency." Both lines are self-reported; neither card publishes benchmark tables underneath them.

To get an MoE running on a phone, the team layers three tricks: Lookahead Expert Scheduling to preload the next layer's experts, fine-grained layer-wise quantization, and a DSpark draft model for speculative decoding. The architecture also uses intra-ViT early compression from LLaVA-UHD v4 with mixed 4x/16x visual token compression, letting developers trade visual detail for inference cost.

The release is Apache 2.0 and ships with FP8, AWQ, GPTQ and GGUF quants, with "Day-0 support for chip platforms: Rockchip, Houmo, Intel, AMD, and Arm," plus iOS, Android and HarmonyOS. Two researchers we follow shared the collection within a day of the drop.