How the X
algorithm works

The 2023 open source (twitter/the-algorithm) described three stages: candidate sourcing (~1,500 posts), ranking with a Light Ranker then a Heavy Ranker neural net, and heuristics & filters. The 2026 repo (xai-org/x-algorithm) keeps Home Mixer as the orchestrator but replaces the hand-built machinery: in xAI's words, they “eliminated every single hand-engineered feature and most heuristics from the system.” Four components do the work now: Home Mixer (orchestration, Rust), Thunder (in-memory in-network post store, Rust), Phoenix (the Grok-based transformer that scores, JAX/Python with a Rust serving layer), and the Candidate Pipeline framework (Rust) everything is composed from.

Thunder holds recent posts in memory as they're published and returns those from accounts you follow (config allows up to ~1,200). Phoenix retrieval is the two-tower model: it embeds the viewer and each post as vectors and returns the nearest posts (up to ~1,000) — pure out-of-network discovery. SimClusters groups accounts and posts by who engages with what and mines the clusters for more. Candidates are then hydrated — text, media, author details, labels, engagement counts — and pre-scoring filters strip out duplicates across sources, posts older than 48 hours, your own posts, blocked/muted accounts and keywords, already-seen posts, and subscriber-only posts you can't access.

Three scorer stages run in order: PhoenixScorer emits a probability per action; RankingScorer takes the weighted sum, then applies author diversity (each post after an author's first is multiplied by a decaying factor, down to a floor), an out-of-network discount (a factor below 1 for accounts you don't follow), and a new-author boost (under ~1,000 impressions, lifted toward slots ~15–16). VMRanker then calls a separate service that reorders with a determinantal point process over post embeddings — trading a little score for less similarity between neighbors, so your feed isn't five versions of the same post. Then TopKScoreSelector keeps the top K. The single most misunderstood line in the repo: weights scale your predicted probability of each action — driven substantially by your own behavior — not raw engagement counts. A report's weight being 468× a like's does not mean one report cancels 468 likes.

Labels are produced continuously, off the request path: classifiers on text and media (grox, media models), account models like agatha (blocks and reports relative to favorites) and bdsm (inauthentic behavior over time), user-cred-v2 (PageRank over the follow and engagement graph), and rule engines (scarecrow embedding botmaker). On the request path, the visibility rules read those labels plus your own blocks, mutes, follows, and settings. A post dropped post-selection takes its thread, quote, or repost ancestors down with it. And you can audit some of this yourself: the Under the Hood page in Settings lets you download a JSON of the labels applied to your account and posts.