In this paper, we propose a dynamic weight tuning scheme for online mu- sic recommendation. Based on a latent factor model, songs, artists, and users are mapped into a latent space. Then, given each user’s recent songs we can determine his current interest for music, which either similar to his past be- havior or more like recent ones. Like latent factor based models, this scheme can be trained without content information, which is a benefit when adopting internet radios as data source. Experimental results on the last.fm collections show that our proposed method is effective. Keywords: Music Recommendation, Dynamic Interests, Latent Factor Model, Machine Learning, Gradient Ascent