X recommendation algorithm

For each request, Twitter said in 2023, it pulls about 1,500 candidate posts from people you follow (in-network) and people you don't (out-of-network) ¹. On average the timeline was half and half, though this varied by user ², and the in-network search index supplied about 50% of posts ³. The 2026 code keeps the split: recent posts from followed accounts, plus posts found by ML retrieval for accounts you don't follow ⁴.

In 2023 the heavy ranker was a neural network of about 48 million parameters ¹ whose ten outputs were each the probability of an engagement ². The 2026 code ranks with a transformer and adds the predictions up as weight times probability ³⁴. A reply is weighted 5.0 against 0.5 for a like ⁵⁶, and X notes the weights multiply your own predicted probability, not raw counts ⁷. Posts from accounts you don't follow are then multiplied by a factor below 1 ⁸.

Before scoring, the 2026 pipeline drops duplicates, posts older than 48 hours and the viewer's own posts ¹. Whether a post may be shown at all is a separate decision, made by visibility filtering ². Twitter first released its recommendation code in March 2023 ³. xAI's newer repository adds the weights that blend predictions into a score ⁴, while some spam rules stay private to make the system harder to game ⁵.