Nowadays, training a well performed deep neural network on CIFAR10 dataset is not difficult problem anymore, but it’s still a difficult problem for neural network models to converge their best training result within a limited time. Current state-of-art deep learning models usually have hundreds of convolutional layers, training epochs and a low learning rate to achieve the best performance which around 99.5% accuracy. Even though those methods can good result, researchers still need to spend long time to training those models on expensive GPU which is uneco-nomic. Therefore, how to balance the training time and best accuracy is always a good deep learn topic to be re-searched. In this paper, I describe a new deep learning model based on Resnet neural network to classify CIFAR10 dataset. This deep learning model has 34 layers including 8 convolutional layers, 8 Batchnorm layers,10 Relu activation layer, 4 pooling layers, 3 linear layers and 1 dropout layer. Within 10 minutes, my neural network model achieved 87.6% accuracy which is better than VGG16, Resnet101 and my assignment2 models and it has potential to achieve higher accuracy with longer training period