<p>Scene text recognition (STR) has demonstrated impressive performance, comparable to human visual capabilities. However, existing methods typically encounter a trade-off, achieving either high accuracy with significant model complexity or lightweight designs that compromise accuracy. There exists a pressing need for supercomputing methods, as current high-accuracy approaches typically demand excessive parameters and prohibitively long inference times, rendering them unsuitable for future real-time deployment on mobile devices. In this paper, we present a Dual-Stream Recursive network based on SVMamba (DSRN-SVMamba). It achieves an optimal balance between accuracy and computational efficiency, utilizing a moderate number of parameters. Inspired by VMamba which not only has powerful feature extraction but also grows computational complexity linearly, we simplify it as our backbone which named SVMamba. Specifically, we reduce the number of Visual State-Space (VSS) blocks in VMamba making it lightweight. Then, we build a dual-stream network consisting of Convolutional Bottleneck Attention Module (CBAM) and Bidirectional Gated Recurrent Unit (BiGRU) after SVMamba. The first stream, CBAM, fully captures the key information in features while maintaining low computational cost and better speed. In the second stream, we utilize the characteristics of BiGRU to collect rich contextual information in sequence data. Additionally, recursive learning is employed in dual-stream network to further limit the parameter count, boosting the practicality and efficiency of the model. Extensive experiments on widely used STR benchmarks demonstrate that DSRN-SVMamba exhibits supercomputing capabilities, achieving a well-rounded performance. The relevant code is available at <a href="https://github.com/jinyuemei/DSRN-SVMamba.git">https://github.com/jinyuemei/DSRN-SVMamba.git</a></p>