OpenBMB releases open weights for MiniCPM-V 4.7 35B-A3B, the series' first MoE vision-language model (community reports)

Verified (fact)

OpenBMB has released MiniCPM-V 4.7 on Hugging Face. It is an MoE model with 35B total parameters, of which about 3B are active per token, and it is the first MoE model in the MiniCPM-V series. GGUF versions were released alongside it. The model supports single-image, multi-image, and video understanding, and uses a technique called Lookahead Expert Scheduling (LES) to target on-device inference.

Needs fact-check (claim)

OpenBMB describes the model as the first open-source MoE model to run smoothly and losslessly for on-device inference (quoting the model card, per Reddit).

Why it matters (interpretation)

This is an attempt to run an MoE multimodal model with about 3B active parameters directly on devices. Developers looking to build image and video understanding locally or at the edge now have a new option.

Sources

• Community openbmb/MiniCPM-V-4.7-35B-A3B Reddit