MobileOne

Apple's high-performance backbone achieves sub-1ms inference on iPhone 12 while maintaining 75.9% ImageNet top-1 accuracy.

MobileOne redefines on-device efficiency by using structural re-parameterization to decouple training-time complexity from inference-time speed. During deployment, the architecture collapses into a linear chain of standard 3x3 convolutions (zeroing out branch overhead). It clocks 10x faster than MobileViT on mobile hardware and outperforms EfficientNet-B0 in both latency and precision. This is the go-to choice for developers needing real-time computer vision without draining the battery or sacrificing model depth.

https://github.com/apple/ml-mobileone