The first MoE model in the MiniCPM-V line, and the point of it is where the sparsity is spent: a 35B total / 3B active vision-language model that runs at 13+ tokens/s in about 3 GB on an iPhone 17 Pro Max. MiniCPM-V has always been an on-device series, so moving to MoE is less about headline capability than about buying a 35B parameter budget at a 3B activation cost on hardware that cannot hold a dense 35B at all. OpenBMB reports a higher overall score in less inference time than Qwen3.6 35B-A3B in Instruct mode.

The enabling piece is Lookahead Expert Scheduling, which addresses the problem that makes MoE awkward on a phone: experts must be paged in per token, and the memory traffic rather than the FLOPs becomes the bottleneck. Paired with fine-grained layer-wise quantization and shipped across BF16, FP8, INT4, AWQ, GPTQ and GGUF builds, plus a 1B sibling for the smallest targets. Apache 2.0.

It is a derivative build, not a from-scratch model: language backbones are Qwen3.5 and Qwen3.6, with SigLIP2 as the vision encoder, so the parameter count belongs to Alibaba's base and is recorded as such. OpenBMB's contribution is the multimodal post-training and the on-device MoE serving path. No paper at filing — this shipped weights-first, which is the pattern the HuggingFace org probe exists to catch.

Model Details

Architecture MOE
License Apache 2.0
Base model qwen3.6-open

Variants

Name Parameters Notes
MiniCPM-V-4.7-35B-A3B — 35B total / 3B active MoE; BF16, FP8, INT4, AWQ, GPTQ, GGUF builds
MiniCPM-V-4.7-1B — smallest on-device target
multimodalmoeefficiencyedgeopen-weights

Related