In recent years, scene text detection has witnessed rapid progress especially with the recent development of convolutional neural networks. However, there still exist many challenges in applying very deep networks to many real-world applications, that have hardware limitations, such as robots, and smartphones. To address these challenges, in this paper, we propose the OctShuffleMLT, an effective fully convolutional neural network, with fewer layers and parameters, which can precisely detect multilingual scene text. Our proposed model is based on the Octave Convolutions that use compact blocks, which reduces memory inference by 13.16%, FLOPS by 71.86%, and the number of parameters by 34.04% when compared to the baseline system. Extensive experiments were conducted on ICDAR 2015 and ICDAR 2017 datasets. Experimental results show that our model can produce accurate detection recognition results on both datasets. The code for the paper is made available on the GitHub repository https://github.com/victoic/OctShuffle-MLT.