ModelBest Open-Sources MiniCPM-V 4.7, Its First MoE Multimodal Model, Running in 3GB on a Phone
ModelBest's MiniCPM-V 4.7 comes in 35B-A3B and 1B sizes. The MoE version decodes at over 13 tokens/s in about 3GB of memory on an iPhone 17 Pro Max, the company says.
ModelBest open-sources MiniCPM-V 4.7, its first mixture-of-experts (MoE) multimodal model, optimized for efficient on-device execution on mobile devices like the iPhone 17 Pro Max.
What happened
ModelBest released MiniCPM-V 4.7, an open-source multimodal model featuring MoE architecture. The 35B-A3B variant claims to run efficiently on mobile devices with minimal memory usage, while the 1B version improves upon previous iterations. The release includes technical innovations like LES (lookahead expert scheduling) and SRFT (Self-Replay Fine-Tuning) for optimization.
Why it matters
Efficient on-device AI models reduce reliance on cloud infrastructure, enhance privacy, and enable real-time processing for mobile users. The open-source nature and cross-platform support lower barriers for developers to adopt and customize the technology.
- On-device AI is becoming more powerful and efficient, reducing cloud dependency.
- Open-source models with MoE architectures offer scalable performance for diverse hardware.
- Innovations like LES and SRFT address key challenges in deploying AI on resource-constrained devices.
The bigger picture
The model enables advanced AI capabilities on mobile devices, potentially transforming applications like offline document processing and private data analysis. Its open-source framework encourages broader adoption and customization in the AI community.
ModelBest's MiniCPM-V 4.7 advances on-device AI by combining MoE efficiency with open-source accessibility, offering developers and users a robust tool for mobile and local computing tasks.
Technical details
- Mixture-of-Experts (MoE) architecture: Activates ~3 billion parameters per token during inference.
- LES (Lookahead Expert Scheduling): A novel inference acceleration technique for memory efficiency.
- SRFT (Self-Replay Fine-Tuning): A low-cost method for adding new capabilities without significant knowledge loss.
- Supported platforms: iOS, Android, HarmonyOS; compatible with Intel, AMD, Arm, Rockchip, and Houmo AI chips.
FAQ
What makes MiniCPM-V 4.7 unique compared to previous models?
It is the first MoE multimodal model in the series, optimized for on-device execution with reduced memory usage via LES (lookahead expert scheduling).
How does the 35B-A3B model achieve high performance on a phone?
By activating ~3 billion parameters per token and using LES, which accelerates inference while maintaining low memory consumption (3GB on an iPhone 17 Pro Max).
What are the key benefits of open-sourcing this model?
It allows developers to customize and extend the model, fosters collaboration, and lowers barriers for deploying AI on mobile and edge devices.
What benchmarks did MiniCPM-V 4.7 achieve?
It scored 91.72 on MMBench, 91.2 on OCRBench, and 90.21 on CMMLU, outperforming Qwen3.6 35B-A3B in some metrics.
What is SRFT, and why is it important?
SRFT (Self-Replay Fine-Tuning) is a technique to add new capabilities to the model with minimal disruption to existing skills, enabling cost-effective customization.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by Pandaily on Oct 14, 2026.
Comments
Sign in to join the discussion.