How does the X algorithm rank a launch post?
What X’s own open-source code and 2026 engineering notes actually say ranks a post, what is still guesswork, and what it means for a launch.
Madni Aghadi · 4 min readX has published two rankers: the 2023 "heavy ranker," a 48-million-parameter model with weights showing replies worth far more than likes, and the 2026 "Phoenix" model, a transformer that predicts 19+ actions but keeps its weight values private. Both treat replies, especially author-to-reply replies, as the strongest documented signal. Links and account coordination are policy questions, not fixed code rules.
Most answers to "how does the X algorithm work" are screenshots of a 2023 config file next to guesses about what changed. X has published two different rankers since then, which agree on some points and disagree on others.
What did X actually open source, and when?
Twice. The first, March 2023, was the "heavy ranker," a 48-million-parameter neural net, published as twitter/the-algorithm and twitter/the-algorithm-ml, covered by TechCrunch.
The second, xai-org/x-algorithm, is newer and still changing: xAI published it January 20, 2026, per ppc.land, then shipped its largest update May 15, 2026, per TechTimes: Phoenix, a transformer trained in JAX with a Rust serving layer, under Apache 2.0, replacing the heavy ranker.
How did the 2023 heavy ranker score a post?
Score equals the sum, over every predicted action, of that action's weight times the predicted probability of it happening. The the-algorithm-ml README lists the weights as they stood on April 5, 2023:
| Signal | Weight (April 2023) |
|---|---|
| Reply engaged by author | 75.0 |
| Reply | 13.5 |
| Good profile click | 12.0 |
| Good click | 11.0 |
| Good click v2 | 10.0 |
| Retweet | 1.0 |
| Favorite (like) | 0.5 |
| Video playback 50%+ | 0.005 |
| Negative feedback v2 | -74.0 |
| Report | -369.0 |
Source: twitter/the-algorithm-ml README, dated April 5, 2023, which says the weights were "originally set so that, on average, each weighted engagement probability contributes a near-equal amount to the score," then adjusted over time. Still open: a GitHub issue asking whether a quote post counts as a retweet here.
Per TechCrunch, roughly half a timeline came from followed accounts, half from accounts a user doesn't follow, after candidate sourcing, the heavy ranker, and filters cut the pool down.
What changed with the 2026 Phoenix model?
X replaced the heavy ranker with a single transformer and stopped publishing weight values. The xai-org/x-algorithm README uses the same shape of formula but a longer, more granular action list: engagement (favorite, reply, repost, quote, share, DM share, copy-link share), clicks (post, profile, link, photo, video), attention (video quality view, dwell time, active seconds), follow author, and negative signals (not interested, mute, block, report).
TechTimes reported xAI redacted the weight numbers "citing security reasons." Viral claims repeating 2023 numbers against this model misapply three-year-old data.
Three adjustments don't appear in 2023: an author diversity decay discounting repeat posts from one author in a feed load; an out-of-network discount on posts from accounts the viewer doesn't follow; and a new-author boost lifting low-impression accounts. No published numbers, just the mechanism.
Do links in posts still get penalized?
Officially, not as of mid-2026, but this has moved before. From late 2022, X got more aggressive about posts sending readers off-platform, and Musk acknowledged in 2024 that linked posts got reduced reach. In July 2026, asked directly, Musk answered "we haven't for over a year," per ppc.land and techtrendske. Musk frames this as a side effect of maximizing time on X, not a hardcoded rule, so it can flip again without announcement.
Is it against X's rules to book several creators to post in the same hour?
No, as long as every post is original and disclosed. X's authenticity policy prohibits "operating multiple accounts that post substantially similar or identical content to one another," plus coordinated activity meant to artificially amplify a topic. It doesn't prohibit several real people posting their own take on the same news at once. The line is content, not timing: identical lines look like coordination; different, disclosed takes are a launch.
Looks like manipulation
- Same sentence copy-pasted across accounts
- No paid-partnership disclosure
- Accounts with no history posting only about this launch
- Engagement that looks automated: identical timing down to the second
Looks like a launch
- Each creator writes their own post
- Every paid post disclosed
- Established accounts with a normal posting history
- Posts land in the same rough window, not the same instant
What's the single best-documented lever for a launch post?
Replies, specifically ones the poster replies back to. Both versions of the code track this as a distinct, heavily weighted signal, more consistently than any other lever here. Everything else (links, quote-post weight, several creators posting together) is a policy question or an unpublished number.
None of this is a formula to plug a launch into. It's what the code and X's own statements support, next to the much longer list of numbers circulating online that neither repo backs up.
Want the full run sheet, not just the ranking mechanics? Read the coordinated launch playbook, browse real launches on X with their numbers, or start a campaign.
Questions people ask
No, not confirmed. Those exact numbers are dated April 5, 2023, from the first open-source release. The 2026 Phoenix model replaced that ranker and xAI has not republished matching weight values.
In the 2023 code, yes, by a wide margin: a reply the author replies back to carried a weight of 75 against 0.5 for a like. The 2026 code still scores reply and author-reply-back as separate signals, but has not published its current weight values.
X applied reduced reach to posts with external links from late 2022 into 2024. Elon Musk said in mid-2026 that X had stopped this "for over a year." Treat this as a live setting that can change again, not a fixed rule.
Not if each account posts its own original, disclosed content. X’s authenticity policy bans accounts posting identical or substantially similar content to each other, not multiple real people posting their own genuine takes around the same time.
Replies, specifically ones the original poster replies back to. Both the 2023 and 2026 versions of the code track this as a distinct, heavily weighted signal, more consistently than any other lever discussed here.
Plan your launch on Goated
See every creator’s rate before you commit, then send the live link to all of them at once.
How it works