X’s recommendation algorithm, open-sourced on January 20, 2026, uses a component-based pipeline called CandidatePipeline. This modular design allows different components to work together seamlessly. The system comprises two main sources for fetching posts: Thunder, which retrieves content from accounts users follow, and Phoenix, a discovery engine that identifies relevant posts based on user interests through machine learning. Thunder focuses on in-network posts, while Phoenix handles out-of-network discoveries, with a scoring penalty for the latter during the ranking process. The pipeline operates through several stages. First, it narrows down millions of posts to thousands of candidates. Thunder maintains a real-time index of recent tweets, while Phoenix employs a two-tower neural network for retrieval, processing user engagement history and candidate features. After identifying candidates, the system classifies them as in-network or out-of-network, ensuring that posts from followed accounts are prioritized. Pre-filtering then eliminates duplicates, incomplete data, and ineligible content, leaving only unique and fresh candidates for scoring. In the scoring phase, each post is evaluated based on predicted engagement metrics, including likes, replies, and shares. The algorithm generates a single score for each candidate, ensuring that scores are independent of batch composition. To maintain diversity in the feed, X applies an exponential decay function to limit the influence of prolific authors. Finally, a selection process chooses the top posts, applying visibility filters to ensure compliance with safety standards before returning the final candidates to users. This meticulous approach balances relevance with diversity, optimizing user engagement while managing content safety.