Ranking model in Douyin’s recommendation system entered the era of one-billion-parameter scale.

In 2025, the ranking model in Douyin’s recommendation system entered the era of one-billion-parameter scale. By introducing the RankMixer ranking model, system computing utilization rose sharply from 4.5% to more than 45%, while inference core activity increased to 80%. At the same time, through algorithmic innovation, the technical team significantly improved training throughput while reducing bandwidth requirements by 84%, successfully enabling ultra-long behavior sequence modeling with lengths of up to 10,000. This breakthrough supports a more comprehensive and deeper understanding of user interests and helps diversify recommended content.