A large scale word-embedding based implicit music recommender is proposed to address the problems that most of current recommendation systems cannot work in the scenario of large scale implicit feedback recommendation. This model employs the Word2Vec technique which is popular in Natural Language Processing in recent years. By learning the songs co-occurrences in the users' history collections, we can get the distributed representation of users and songs with a low-dimension and dense vector. In this way, we can get the similarities of users and songs which could be used for the recommendation and we also analyze the correctness of application of Word2Vec technique in recommendation. This model can effectively solve the problem mentioned above with the accuracy remaining the same. In addition, this model can converge faster and take less memory than those of traditional methods.