Placeholder Repository for MiniCPM-V-4.7-35B-A3B Spotted on Hugging Face
A placeholder repository for an upcoming multimodal language model, MiniCPM-V-4.7-35B-A3B, has appeared on Hugging Face. The model originates from OpenBMB's MiniCPM-V series, though official weights and model details have not yet been populated. The MiniCPM-V lineup is widely recognized in the open-source community for delivering strong multimodal visual-language understanding with high parameter efficiency. Scaling the architecture to a 35B total model with 3B active parameters suggests OpenBMB is preparing a much more capable model while attempting to maintain low compute requirements. The model repository currently lacks a completed model card, benchmark scores, architecture breakdown, or downloadable weights. The identifier '35B-A3B' implies a sparse Mixture-of-Experts (MoE) design containing roughly 35 billion total parameters with only 3 billion active parameters per forward pass.
## BACKGROUND
MiniCPM-V is a series of open-source multimodal large language models developed by OpenBMB focused on vision-language tasks across image, video, and text inputs. Earlier versions, such as MiniCPM-V 2.6, paired vision encoders like SigLIP with compact base language models like Qwen2 to achieve efficient performance suitable for edge device deployment.