How the X algorithm works in 2026, according to its own code

Where the 2026 algorithm comes from

In January 2026 X published the code behind its For You feed on GitHub as xai-org/x-algorithm, under the Apache 2.0 licence; the first commit is dated 20 January. It follows twitter/the-algorithm, which Twitter released in March 2023. That older system ranked posts with a neural network called the heavy ranker, fed with features describing each post and user. The January 2026 README said X had “eliminated every single hand-engineered feature and most heuristics” and ranked posts with a “Grok-based transformer”.

It is not a one-off release. Since August 2026 the repository has been updated several times a week, and its README says X runs scripts that set the defaults in the code to the primary production values, and aims to show experiments running on a notable share of traffic, such as 10% or more. The weights file read for this guide says it was last synced on 7 October 2026.

How a post reaches your For You feed

The README describes a feed built fresh for every request, in this order:

1. Your context. The system loads your recent engagements, which the README calls “the main input to the model”, plus who you follow, your blocks and mutes, muted keywords and the posts you have already seen.
2. Candidates. Thunder holds recent posts from accounts you follow. For accounts you don't follow, Phoenix retrieval returns posts whose embeddings sit nearest yours, SimClusters finds posts through clusters of who engages with what, and a shared list adds popular recent posts.
3. Filters. Before scoring, the code removes duplicates, posts older than 48 hours, your own posts, accounts you block or mute, posts matching your muted keywords, posts you have already been shown, and reposts and replies from accounts you don't follow.
4. Scoring. A model called Phoenix scores every remaining post, and the highest scores are kept.
5. Visibility. A separate system decides whether each kept post may be shown to you at all.
6. Blending. Ads, Who to Follow and prompts are mixed in, the last two at fixed positions.

What the model predicts, and the weights

For every candidate post, Phoenix predicts how likely you are to like, reply, repost, quote or share it (by DM or by copying the link); to click the post, the profile, a link, a photo or a video; to dwell on it or keep watching; to follow the author; or to react negatively with not interested, mute, block, report or not dwelling on it. The scorer adds them up as Σ (weight × P(action)).

Read them the way the comments in that file ask. Each weight multiplies a predicted probability, not a count: the comments call “one report cancels 468 likes” incorrect, and note that a report's baseline probability is more than 1,000 times lower than a like's, which is why its weight is larger. A reply weighted ten times a like is not worth ten likes; what each term adds depends on how likely you, the viewer, are to take that action.

Numbers still quoted around the web, such as 13.5 for a reply, 75 for a reply the author engages with or −369 for a report, are the April 2023 weights of the old heavy ranker, published in twitter/the-algorithm-ml. The 2026 list of predicted actions has no “reply engaged by author” at all.

Three adjustments after scoring

• Author diversity. Each further post from the same author is multiplied by a decaying factor, down to a floor.
• Out-of-network discount. Posts from accounts you don't follow are multiplied by a factor below 1, and so are replies and reposts from accounts you do follow.
• New-author boost. In author_cold_start.rs, with the defaults synced on 7 October 2026, a post qualifies when it is an original post (not a reply or repost), less than two hours old, with fewer than 200 views on Home, from an author with 50,000 followers or fewer. At most one qualifying post per ranked batch is lifted to at least the score of the post in 16th place, picked by a random draw that favours posts with more likes per Home view.

A final step, VMRanker, reorders the list so that similar posts don't sit next to each other, giving up a little score to do it.

What changed in 2026

• January. The For You code goes public as xai-org/x-algorithm, first commit dated 20 January.
• 15 May. An update adds a runnable end-to-end inference pipeline and new components for content understanding, ads and candidate sourcing.
• 10 to 24 July. The mutual-follow boost. An A/B test from 10 July gave a small share of users a boost of 5, 10, 15 or 20 on the reply weight for original posts by people they follow who follow them back. On 13 July a value of 20 went out to many users; on 24 July X lowered it to 15 after experiment results and feedback, for example that some people weren't seeing as much World Cup discussion as they wanted (X's write-up of the change). Announcing it, Nikita Bier wrote that it would “boost visibility of your posts to your mutuals (people who you follow back)”. The default synced on 7 October is still 15.
• 13 August. X publishes the configuration values, including the weights above, plus the visibility-filtering code, the systems that label posts and accounts, Phoenix's training code and SimClusters, and launches a transparency tool called Under the Hood.
• 14 August. Comments explaining how the weights work, and a filter that removes posts from accounts reported to Brazil's Electoral Court for the 2026 election, unless you follow them.
• 18 September. Under the Hood reports start showing whether posts were withheld in a country after a legal demand.

Ranking is not visibility

Ranking sets the order; visibility filtering decides whether a post can be shown at all. For each post and viewer it answers allow, notice, interstitial or drop, using your own blocks and mutes and labels that other systems attach to posts and accounts. Some rules apply only when a post is recommended to someone who doesn't follow its author, “spam caught at high recall, for instance”; the same post is still allowed to a follower.

One labelling job, agatha, labels accounts from how others respond to their posts: blocks, reports and spam reports relative to likes. Not everything is published: the README says some labelling rules and the LLM prompts of its post classifiers are held back to make gaming harder. Instead, X is piloting Under the Hood, which shows the visibility-limiting labels on your own account and posts, with availability to expand over time.

Replies under a post are ranked separately

The open-source code covers the For You feed. Inside a conversation, X's help page on the For you timeline says that when ranking a reply higher it considers “if the original post author has replied, if a reply is from someone you follow, or if the person is a X Premium subscriber”. The Premium page adds that people “will see a slight preference for replies from verified accounts”, that X is still testing how much, and lists reply prioritization on every Premium tier, larger on Premium and largest on Premium+.

In the For You code, replies from accounts the viewer doesn't follow are filtered out (OONRetweetReplyFilter; the switch that would let some through was off in the 7 October defaults). People who don't follow you see your replies inside the conversation, not in their For You feed.

What it means for a small account

The code shows what is scored, not what will work for you. The points below are our reading of it, each tied to the code above.

For a worked example, the founder's paid guide Talk First describes how he went from 0 to 2,699 followers in 34 days.

Where xInsightss fits

xInsightss can't change how X ranks your posts, and nothing in it promises reach or followers. It helps with the parts you control:

• The reply inbox collects every reply to you in one queue, so you can answer while the conversation is fresh.
• Engage finds fresh original posts on your topics, from the last 24 hours, 3 days or 7 days, and opens X with your reply ready to post.
• Analytics judges every post over its first 48 hours against your follower count, so you can see which posts travelled beyond your followers.

Questions