The X algorithm for founders: what the open-source code rewards and how to read Under the Hood

What X has published, and what it has not

The repository xai-org/x-algorithm holds the code that builds the For You feed: where candidate posts come from, the filters, the model that predicts how a viewer will react to each post and the scorer that turns those predictions into a ranking. X created it in January 2026 and expanded it on August 13, 2026 with the weighting parameters. It is updated often; the version read for this guide is commit 2f25214 from October 26, 2026, and its parameter file says it mirrors X's configuration defaults as of October 26.

On October 20, 2026, Elon Musk announced two changes on X: a new ranking model that predicts probable time spent much better, and a retrieval model trained to surface posts people read all the way through. X has not said how time spent is weighed against replies and shares, or whether every account has the new models. At commit 2f25214 the published weights are unchanged, including dwell at 0.05, so read the announcement as better predictions of attention, not a new reward table. Our reading: it points the same way as the weights, toward posts your buyers stop to read and answer.

The code does not show your own scores, the prediction model's output for any post, the experiments X is running on a given day, or the Grok-based systems that predict rule violations, which X excluded on purpose. Treat it as an accurate description of how ranking is designed, not as a readout of why one post did well.

How a post reaches someone's For You feed

For each viewer, the system gathers candidate posts, drops ineligible ones, predicts the viewer's reactions, scores and selects the top posts, and runs visibility filtering again before the feed is shown.

1. Candidates: posts from accounts the viewer follows (in-network) and posts from accounts they do not follow (out-of-network), found by a retrieval model, topic clusters (limited to posts up to 48 hours old) and a shared list of popular posts.
2. Pre-scoring filters: duplicates, posts the viewer has already seen or been served, muted keywords, blocked or muted authors, posts that are too old and the viewer's own posts, among others.
3. Budget: when there are more candidates than the model will score, a fair-share filter (switched on by default between the October 20 and October 26 parameter syncs) keeps the same fraction from each candidate source, so posts from followed accounts and recommendations are trimmed evenly.
4. Prediction: a transformer model estimates the chance that this viewer will like, reply, repost, quote, share, click, dwell, follow you, or say not interested, mute, block or report.
5. Scoring: each prediction is multiplied by its weight and summed. Posts from accounts the viewer does not follow are scaled down by an out-of-network multiplier below 1.
6. Selection and visibility: the top posts are kept, a reranking service orders them, and visibility filtering removes or limits posts with certain labels.

The weights as of October 26, 2026

These are the default weights in home-mixer/params/param.rs. A weight only matters together with the predicted probability it multiplies: a like is predicted far more often than a report, so the large negative weights matter most for posts some viewers are likely to dislike.

Three code paths that matter more than any trick

Negative scores rank last. The scorer maps any post whose weighted total is negative into a range below every post with a positive total. For that viewer, a post that enough people would hide or report starts below every post with a positive total, however many likes it is also predicted to get; only the final reranking step can change that order.

Small accounts get an exploration slot. A cold-start scorer can place a post into a slot near position 15 of the feed when its author has at most 50,000 followers, the post is at most two hours old and it has fewer than 200 impressions. It runs as an experiment with control and holdout groups, so not every viewer sees it, and it gives a new post a chance to be seen, not a guaranteed audience.

Out-of-network reach is discounted. A post shown to someone who does not follow you is multiplied by a factor below 1, so strangers see it only when its predicted reactions are strong. For an early founder, your existing followers and mutuals are where a post earns its first replies.

Worked example: scoring two posts

Everything below is fictional: the product, the posts and every probability. Real scores combine about 30 predictions made by X's model for each viewer, which nobody outside X can see. The example uses five of them to show how the weights interact, not to predict reach.

Ledgerline is an invented bookkeeping tool for agencies. Its founder drafts two posts. Post A announces a launch with a link. Post B asks agency owners how they chase late client payments. For one viewer who follows the founder but not mutually, the model might predict the chances below.

Read your Under the Hood report

Under the Hood is X's page for downloading a report of the visibility labels applied to your account and posts in the prior month. Since September 18, 2026 it also shows whether posts were withheld or limited in a country because of a legal demand. X says it is available to accounts at least one year old that posted 10 or more times in the prior month; open x.com/i/under_the_hood or find it in your account settings. If you see only an explanation and no download, your account is not eligible yet.

Read it for what it can tell you: whether a label limited where your posts could appear. It cannot tell you why a post got few impressions, and reviewers report that it gives counts and percentages per label rather than naming each affected post.

Claims the current code does not support

Many posts about the X algorithm repeat numbers from the 2023 release or from earlier versions of this repository. Check any claim against the commit it came from.

Copy the weekly review

Copy this review into your own document or spreadsheet; no account is required. It keeps the signals the code weights heavily separate from the outcome you actually need: conversations with people who might buy.

Where Soloop fits

Soloop is an AI cofounder that runs individual go-to-market tasks for a founder. According to the Soloop manual, after you connect your X account Soloop can write posts and replies, publish them through the available account controls and track responses. It can also research a market and return the sources it used.

Soloop has no relationship with X, cannot see your Under the Hood report and cannot change how X ranks your posts. A bounded first request: My product is [URL] for [role]. Find recent public X posts where [role] asks about [problem], summarize the questions with sources, and draft two question posts and three replies I could add to those threads. Do not publish until I approve.

Soloop does not promise impressions, replies or followers. Check the current X connection and publishing steps in your workspace, because they can change.

Sources, limitations and refresh

Weights, filters and scoring behavior come from the xai-org/x-algorithm repository at commit 2f25214, read on October 26, 2026. Under the Hood eligibility comes from X Open Source's post on September 18, 2026. Report structure and the August release details are reported by news coverage and third-party guides. The founder meanings, ratios, claim checks, review and fictional example are editorial guidance, not a measured study. Under the Hood eligibility was rechecked on October 26, 2026. The October 20 model change is reported from Musk's post as quoted by Tom's Guide.

The repository mirrors production defaults but not every experiment, and X changes them often. Recheck the weights against the latest commit before relying on them. This guide will be reviewed by January 26, 2027.