MiniCPM-V 4.7 brings a 35B MoE vision model on-device

OpenBMB has released MiniCPM-V 4.7, the newest version of its on-device multimodal line, as a mixture-of-experts vision-language model. The configuration pairs 35 billion total parameters with roughly 3 billion active per token, a design meant to deliver the capacity of a larger model while keeping the per-inference compute cost closer to a small one.

The model is distributed in GGUF format on Hugging Face, the quantized packaging favored by local runtimes like llama.cpp and Ollama. That choice signals the intended audience: developers who want to run image understanding on laptops, workstations, or edge hardware rather than calling a hosted API.

Why it matters

MiniCPM-V has built a following by targeting the practical middle ground between tiny toy models and server-only giants. A sparse MoE approach extends that logic:

• Higher ceiling: 35B total parameters give the model more knowledge to draw on.
• Lower active cost: only about 3B parameters fire per token, easing memory and speed demands.
• Local-first delivery: GGUF builds make it straightforward to run offline.

As with prior MiniCPM releases, the model ships under a custom license, so teams should review the terms before commercial use. OpenBMB has not published full context-length or benchmark details in this record, but the release continues the project's focus on bringing capable multimodal reasoning to hardware people actually own.