This talk will introduce our work on AI based Antivirus using deep learning. We can control the false positive rate less than 0.05% and false negative rate less than 12%. We think it's OK for production and it's already in production since Jan 2016. Android malware can often evade anti-malware security software if the author changes a few lines of code or designs the program to automatically mutate before each new infection or add shell on their app. Deep learning involves training an artificial neural network with many layers of simulated neurons using huge quantities of data. The networks trained to recognize the characteristics of malicious code by looking at ten million of examples of malware and non-malware files, could offer a far better way to catch such malicious code. We build a deep learning system for Android anti-malware. We select high-quality app features data with only a little size, and use innovative normalization preprocessing, unique activation function and advanced multilayer artificial neural network to recognize the unknown malware variants and defense zero-day attacks. Our deep learning system has high precision (99.96%) and high recall (88%).