Airbnb developed a Transformer-based sequence model to personalize listing recommendations by understanding guest behavior over time. Traditional methods relied on handcrafted features, which became difficult to scale and limit expressiveness. The new model addresses challenges like the sheer volume of listing views and the rarity of bookings compared to other events. It splits guest history into a long-term sequence of infrequent but informative events and a short-term sequence of recent listing views. To improve training efficiency, they implemented strategies like batching searches and minimizing padding with bucketing and sparse calculations. The serving design decouples the computationally intensive sequence encoder from the real-time ranking model for low latency. Initial testing with the long-term sequence alone showed improvements, which were further boosted by adding the short-term view sequence. The introduction of a setwise ranker, co-trained with the sequence encoder, led to even greater ranking quality. Online A/B tests confirmed these gains, showing significant increases in bookings and views. The system's effectiveness was also demonstrated in promotional email ranking, highlighting the generalization of guest embeddings. This new approach allows Airbnb to understand guest preferences more deeply than before. Future work includes real-time updates, richer event types, and generative recommenders.