X (Twitter) Open-Source Algorithm: Every Release of the xai-org/x-algorithm Repo, Dated and Explained

Published: 20th Jan 2026 Last updated: 4th Sep 2026

X published its For You ranking code on 20th Jan 2026 as 79 files of Rust. Eighteen releases since have added scoring weights, visibility filters, ads blending and the first merged outside fix.

X's For You ranking is public code. The xai-org/x-algorithm repo appeared on 20th Jan 2026 with 79 files of Rust.

Its three parts: Home Mixer runs the pipeline, Thunder supplies in-network posts, and Phoenix is a Grok-based ranking model.

The second release, in May, made the repo runnable and documented ads blending. The third, on 13th Aug, published the scoring weights and visibility filters.

It also shipped an "Under the Hood" tool that lets accounts download their own visibility labels.

Since mid-August releases have landed almost daily. We read each commit from the GitHub API and log what changed.

On 3rd Sep 2026 X merged its first outside pull request, a two-line fix. Each release below links to its entry and the commit on our Social Platforms tracker.

Rollout log

20th Jan 2026
X open-sources its recommendation algorithm, for the second time in its historyThe xai-org repo landed with 79 files of Rust: Home Mixer orchestration, Thunder for in-network posts and Phoenix, a Grok-based ranking transformer. Its README claims every hand-engineered feature has been eliminated.
15th May 2026
X's second algorithm release makes the repo runnable, and documents ads inside the ranking pipelineA 187-file commit added an inference pipeline, a pre-trained mini Phoenix model, the Grox content-understanding service and an ads-blending module. It was also the first public commit in four months, against a promised four-week cadence.
13th Aug 2026
X publishes its ranking weights and its visibility filters, plus a tool that shows your own labelsThe third release adds scoring weights, visibility systems and trainable Phoenix model code, making the codebase roughly 10 to 15 times larger. A paired "Under the Hood" tool lets accounts with at least 10 posts a month download the visibility labels applied to them.
17th Aug 2026
X adds topic-level diversity to For You, in a release it did not announceA fourth commit in five days gives the For You slate builder repetition counts for a post's topic, not just its author. It also sends author follower counts and an NSFW-author flag to the ranker, and grows the Following feed from 100 to 110 results.
18th Aug 2026
X wires its new topic-diversity signals into the ranker, and chains its two brand-safety generationsA fifth commit in six days sends the semantic-ID repetition counts added on 17th Aug to Phoenix with every candidate. A post Grok's v2 labels have never touched now gets the v1 brand-safety verdict, not a medium-risk rating.
19th Aug 2026
X extends muted keywords to the Following timeline, and makes them match quoted posts and reply threadsA sixth release in seven days adds keyword muting to the Following feed, which never had it, and the filter also reads quoted and parent posts. The For You filter still checks only the post's own text.
20th Aug 2026
X builds AI-generated trend labels into For You posts, with a tap-through to the trendA seventh release in eight days tags up to two posts per timeline with the name of an AI-generated trend they belong to. Unlike existing topic labels this one links to the trend page, and both changes ship behind flags that default to off.
21st Aug 2026
X moves the reply-spam boundary from 60,000 followers to 80,000, and wires 20 ad conversion types into the rankerAn eighth release in nine days raises the follower threshold for reply spam-scoring and adds 20 view-through and in-app ad conversion events. It also starts logging when a searcher rephrases a query.
24th Aug 2026
X adds a fifth brand-safety verdict for its most severe labels, and moves the reply-spam boundary again to 100,000 followersA ninth release, after the first pause since the cluster began, splits the worst safety labels out of "medium risk" into a new high-risk verdict. Ads stay away from both, and the follower threshold for reply spam-scoring rises for the second time in three days.
25th Aug 2026
X switches on the brand-safety verdict it shipped disabled, and grows its Brazilian account list from 665 to 2,328A tenth release flips on the ad brand-safety switch, off since 13th Aug, so Grok's second-generation labels now decide ad placement. The Brazilian electoral filter more than triples, dwell time enters the ranking weights and video quality views drop out.
26th Aug 2026
X starts showing follower counts to the model that judges reply spamThe eleventh release in a fortnight tells the reply-spam classifiers each author's follower count and the original poster's bio. A reply's stored ranking score can no longer be replaced by a higher one, and no ranking parameter moved.
28th Aug 2026
X caps ads by the riskiness of nearby posts, and stored reply scores can now only fallThe twelfth release in a fortnight adds an ads blender budgeted against nearby risk labels, a stricter reply-score ratchet and two missing schema fields. It also names ten court-reported Brazilian accounts X has decided not to filter.
28th Aug 2026
X adds a config switch that can change which ranking-model head serves, and locks its moderation rules under a golden test corpusA thirteenth release threads a RerankerHeadTag value, default zero, through every scoring request. The moderation rules also gained a 1,039-line corpus of expected verdicts future changes must edit.
1st Sep 2026
X moves its country-by-country adult-content gating into a live config, and starts recording ranking scores per experiment bucketA fourteenth release publishes the sixteen countries where X gates adult content as a Rust setting reread every thirty seconds. Score distributions can now be recorded per A/B bucket.
1st Sep 2026
X makes the risk-bucket ads blender the default for the home timeline, and rewrites its moderation rules as two tablesA fifteenth release makes the MultiRiskBlender, added four days earlier, the default home-timeline ads blender. Nine moderation rule files became two tables, and the ranker gained a codebook embedding mode.
2nd Sep 2026
X flips its Phoenix aggregation default to a long-dwell variant, and lets its bot detector file reportsA sixteenth release switches the mirrored default for aggregating Phoenix predictions to a long-dwell variant. Its legacy bot detector can now file proactive reports against posts and profiles.
4th Sep 2026
X merges its first outside pull request, then ships a release that adds an LLM reply-spam classifierX merged its first outside pull request on 3rd Sep, a two-line fix. A seventeenth release followed with an LLM reply-spam classifier.
4th Sep 2026
X tightens its cold-start slot for the Phoenix retrieval experiment, and wires per-request overrides into the ranking callAn eighteenth release changes which low-impression posts qualify for the cold-start slot in one experiment arm. The experiment-overrides map reaches the Phoenix request format.

Sources

Frequently asked questions

Is the X algorithm really open source?

The For You ranking code is public under Apache 2.0 at github.com/xai-org/x-algorithm, since 20th Jan 2026. Full trained model weights are not; only a mini demo model is.

How often does X update the public algorithm?

Four months passed between the first and second releases. Since 13th Aug 2026 releases have landed every few days; every one is logged on this page.

Compiled by Claude from primary sources, checked by Akash Mansukhani before publication. See how we source and correct this page.

Spotted a change we have not covered? Ranking volatility, a new SERP test, a platform change: email us a tip and we will credit you if we cover it.

Cite this page:

Keywords Everywhere, "X's open-source algorithm: every release since January 2026", last updated 4th Sep 2026, /kw/news/social-algorithm-updates/x-algorithm-open-source/

X's open-source algorithm: every release since January 2026