OFDM systems have gained widespread application in wireless communication due to its capability to efficiently alleviate multipath fading and mitigate inter-symbol interference. In OFDM systems, accurate channel state information (CSI) is significant to recover symbols from the received signals. In this paper, we propose a multi-head attention mechanism and feedforward neural network (MAF) neural network model, which is mainly divided into encoder and decoder parts. The encoder introduces a multi-head attention mechanism to capture different aspects of the input sequence’s information while a residual convolutional neural network is introduced to address the gradient vanishing and exploding problems in the training process. Additionally, we also introduce a feedforward neural network in the decoder to better handle models with complex structures and different types of data. Simulation results show that the proposed model performs well in channel estimation, exhibiting superior performance compared to other neural network methods. Furthermore, even with a pruning rate \(\le \)  60%, the model maintains robust performance, reducing model complexity and computational costs.