Neural networks have been extensively utilized in various fields, steganography for neural network is a research emerging direction in academia in recent years. Embedding capacity and robustness are important indicators for steganography. But balancing embedding capacity and robustness is challenging. This paper proposes a robust steganography for neural network models. Embedding secret data into neural network without visibly reducing the performance of the original task. This is achieved by embedding secret data during the training process instead of modifying the network parameters after training. Receivers can obtain the secret data from data decoding networks, the parameters of data decoding networks are generated using the embedding keys. In this way, it is unnecessary to transmit the decoding networks secretly. Additionally, introducing Reed-Solomon codes (RS codes) to improve data extraction robustness. Experimental results reveal that the robust steganography for neural models improves robustness while maintaining superior embedding capacity.