Transformer architectures have dramatically influenced the field of natural language processing and are becoming more popular in the computer vision field, too. However, the Transformer’s core self-attention mechanism has quadratic computational complexity concerning the number of tokens. Thus, they usually require big GPUs for deployment, contrary to the Internet of Things trend, which enables the mobile deployment of AI applications, which involves the development of efficient, lightweight neural networks to meet the strict hardware limitations of the target platforms. The cost and ease of deployment are even more critical in the medical field, and not every clinic can afford to buy a lot of powerful GPUs to aid the physicians. Therefore, research proposed some methods to achieve more efficient transformer networks, but to our knowledge, very limited work targeted a level of complexity reduction that allows the embedded deployment of transformers in the medical field. In this paper, we propose our Embedded-ViT framework with which we can drastically reduce the complexity of standard vision transformer (ViT) networks. We accomplish that using several compression techniques: efficient model architecture changes, reduced input resolution, pruning, or quantization. Our optimizations can significantly compress the model while maintaining a desired prediction quality level. We prove the capabilities of our framework by applying it to a state-of-the-art ViT and its variations. We tested the results of our Embedded-ViT on the publicly available Synapse dataset for multi-organ segmentation. Our framework cuts the computational load by half while maintaining a slightly higher level of prediction quality. Moreover, we will thoroughly analyze the hardware requirements and throughput achieved on different platforms, including the embedded Jetson Nano from Nvidia. The framework is open-source and accessible online at https://github.com/ErikOstrowski/Embedded-ViT .