MobileNet is a family of convolutional neural network (CNN) architectures developed for image classification, object detection, and other computer vision tasks. The models are engineered for small size, low latency, and low power consumption, making them suitable for on-device inference and edge computing on resource-constrained devices such as mobile phones and embedded systems. Originally, they were designed to run efficiently on mobile devices using TensorFlow Lite, a lightweight runtime for machine learning models.
The need for efficient deep learning models on mobile devices led researchers at Google to develop MobileNet. As of June 2025, the family includes five major versions, each building upon its predecessor to improve performance and efficiency while maintaining suitability for constrained environments.
Architecture Principles
MobileNet architectures focus on reducing computational cost through several key techniques. The core innovation is the use of depthwise separable convolutions, which decompose a standard convolution into two separate operations: a depthwise convolution that filters each input channel independently, and a pointwise convolution (a 1×1 convolution) that combines the outputs. This factorization significantly reduces the number of parameters and floating-point operations compared to standard convolutions, enabling efficient inference on devices with limited processing power.
Another significant strategy is the introduction of hyperparameters that control model size and computational load. The width multiplier (denoted as α) adjusts the number of channels in each layer; smaller α values produce smaller and faster models at the cost of accuracy. The resolution multiplier (ρ) adjusts the input image resolution, with lower resolutions speeding up processing but potentially reducing precision.
The architecture also incorporates techniques such as batch normalization and data augmentation to improve training stability and generalization. These design choices have made MobileNet a popular choice for real-time applications in mobile and embedded contexts.
MobileNetV1
MobileNetV1 was published in April 2017. Its main architectural innovation was the incorporation of depthwise separable convolutions, a concept first introduced by Laurent Sifre during an internship at Google Brain in 2013 as a variation on AlexNet intended to improve convergence speed and reduce model size. The depthwise separable convolution decomposes each standard convolution into a depthwise filter (operating per channel) and a pointwise combination (1×1 convolution) that merges outputs from the depthwise layer.
Defaults for MobileNetV1 included a width multiplier of 1.0 and a resolution multiplier of 224x224. However, by adjusting these multipliers, developers could balance between form factor and accuracy. The architecture also used ReLU activations after each convolution and applied Batch Normalization to accelerate training.
MobileNetV2
MobileNetV2, published in March 2019, introduced two important refinements: inverted residual layers and linear bottlenecks. Inverted residuals flip the structure of traditional residual blocks by first expanding the input channels, then performing depthwise convolution, and finally projecting back down to a lower number of channels. This expansion allows the depthwise convolution to operate in a higher-dimensional space, preserving more information flow.
Linear bottlenecks remove the ReLU activation from the projection layers, based on the intuition that nonlinearities in low-dimensional spaces cause information loss. By keeping these layers linear, the model retains more expressive power even when channel numbers are small. These improvements yielded higher accuracy and lower latency than V1, making V2 a common choice for mobile vision tasks.
MobileNetV3 and V4
MobileNetV3, released in 2019, introduced three variants: MobileNetV3-Large, MobileNetV3-Small, and MobileNetEdgeTPU, the latter optimized for Pixel 4 smartphones. These models were discovered through neural architecture search (NAS) that directly measured mobile latency, achieving a desired trade-off between accuracy and inference speed. V3 also included piecewise-linear approximations of swish and sigmoid activations (h-swish and h-sigmoid), squeeze-and-excitation modules, and inherited the inverted bottleneck structure from V2.
MobileNetV4, published in September 2024, expanded the family with numerous architectures also found by NAS. V4 introduced multi-query attention, inspired by vision transformers, and proposed a new post-mix block called the universal inverted bottleneck, which unifies the inverted residual and inverted bottleneck concepts from earlier versions.
MobileNetV5 and Future Directions
The architecture of MobileNetV5 was announced shortly after the release of Gemma 3n (a language model) in June 2025. Although the announcement mentioned that a technical report would be published soon, as of October 2025, no such report had been made available. The V5 network is reported to be roughly ten times larger than the largest V4 variant, suggesting a shift toward higher-capacity models while maintaining achievable efficiency for mobile deployment.
Throughout its development, MobileNet has influenced many architectures and remains a cornerstone for embedded vision systems, enabling practical applications in fields like autonomous driving, robotics, and mobile photography.