Early alert fire and smoke detection systems are crucial for management decision making as daily and security operations. One of the new approaches to the problem is the use of images to perform the detection. Fire and smoke recognition from visual scenes is a demanding task due to the high variance of color and texture. In recent years, several fire-recognition approaches based on deep learning methods have been proposed to overcome this problem. Nevertheless, many developments have been focused on surpassing previous state-of-the-art model's accuracy, regardless of the computational resources needed to execute the model. In this work, is studied the trade-off between accuracy and complexity of the inverted residual block and the octave convolution techniques, which reduces the model's size and computation requirements. The literature suggests that those techniques work well by themselves, and in this research was demonstrated that combined, it achieves a better trade-off. We proposed the KutralNext architecture, an efficient model with reduced number of layers and computacional resources for singleand multi-label fire and smoke recognition tasks. Additionally, a more efficient KutralNext+ model improved with novel techniques, achieved an 84.36% average test accuracy in FireNet, FiSmo, and FiSmoA fire datasets. For the KutralSmoke and FiSmo fire and smoke datasets attained an 81.53% average test accuracy. Furthermore, state-of-the-art fire and smoke recognition model considered, FireDetection, KutralNext uses 59% fewer parameters, and KutralNext+ requires 97% fewer flops and is 4x faster.