OpenBMB Releases MiniCPM-V 4.7: 35B Multimodal Model Runs on Mobile with ~3GB Memory

OpenBMB has released MiniCPM-V 4.7, a 35B-A3B MoE (Mixture of Experts) vision-language model designed to run locally on mobile devices with approximately 3GB of memory. This release marks a significant step toward capable on-device multimodal AI, allowing for tasks such as reading real-world signs, understanding complex documents, and reasoning over visual information without cloud dependency.

Key Technical Innovations

• LES (Lookahead Expert Scheduling): Instead of simply reducing model size, this technique optimizes MoE inference by predicting and preloading experts, thereby reducing computational bottlenecks and enabling efficient execution on constrained hardware.
• Self-Replay Fine-Tuning (SRFT): A new method allowing developers to customize the model for specific tasks while preserving its general capabilities.
• MiniCPM-V 4.7 1B: A smaller variant released for devices with even more limited resources.

The model supports planning with real-world constraints and demonstrates strong performance in visual reasoning tasks, offering a viable alternative to cloud-based multimodal services for mobile applications.