OpenBMB's MiniCPM-V 4.7 brings MoE to vision models

OpenBMB has released MiniCPM-V-4.7-35B-A3B, the latest entry in its MiniCPM-V line of multimodal models. The headline detail is architectural: this is a mixture-of-experts (MoE) vision-language model with 35 billion total parameters but only about 3 billion active on any given forward pass.

That design is the whole point. MoE lets a model hold a large pool of specialized parameters while routing each input through a small fraction of them, keeping inference costs closer to a 3B dense model than a 35B one. For a vision-language system, where image understanding typically adds significant compute, that efficiency can be the difference between a model that runs comfortably on modest hardware and one that doesn't.

Why it matters

The MiniCPM-V series has built a reputation for packing strong multimodal performance into compact footprints, and version 4.7 extends that philosophy with an MoE twist. Key points from the release:

• 35B total parameters with roughly 3B active per inference
• A vision-language focus, handling image and text inputs together
• Permissive Apache 2.0 licensing, allowing commercial use

The permissive license is notable for teams weighing deployment. Apache 2.0 removes much of the friction that accompanies more restrictive open-weight releases, making MiniCPM-V 4.7 a candidate for products as well as research.

OpenBMB has not published full benchmark details in the record here, so real-world evaluation will determine how the sparse-activation approach compares to dense peers. But the direction is clear: efficient multimodal models that trade raw size for smart routing. The weights are available now on Hugging Face.