Light-weight convolutional neural networks (CNNs) are the de-facto for mobile\nvision tasks. Their spatial inductive biases allow them to learn\nrepresentations with fewer parameters across different vision tasks. However,\nthese networks are spatially local. To learn global representations,\nself-attention-based vision trans-formers (ViTs) have been adopted. Unlike\nCNNs, ViTs are heavy-weight. In this paper, we ask the following question: is\nit possible to combine the strengths of CNNs and ViTs to build a light-weight\nand low latency network for mobile vision tasks? Towards this end, we introduce\nMobileViT, a light-weight and general-purpose vision transformer for mobile\ndevices. MobileViT presents a different perspective for the global processing\nof information with transformers, i.e., transformers as convolutions. Our\nresults show that MobileViT significantly outperforms CNN- and ViT-based\nnetworks across different tasks and datasets. On the ImageNet-1k dataset,\nMobileViT achieves top-1 accuracy of 78.4% with about 6 million parameters,\nwhich is 3.2% and 6.2% more accurate than MobileNetv3 (CNN-based) and DeIT\n(ViT-based) for a similar number of parameters. On the MS-COCO object detection\ntask, MobileViT is 5.7% more accurate than MobileNetv3 for a similar number of\nparameters.\n Our source code is open-source and available at:\nhttps://github.com/apple/ml-cvnets\n