LinkIntel
Start free trial

Free · no account

X's ranking weights, read from X's own code.

X's algorithm weights are the numbers in X's open-source ranking code that score how much each predicted action counts. In the published defaults, a reply, quote or share via DM is weighted 5, a follow the author 4, a repost 1 and a like 0.708. A report is -234. LinkIntel reads these from X's repo every day and links each value to its commit. These are defaults mirrored from X's open-source repo, not proof of what is live, and not a prediction of reach.

Latest commit590f39213 November 2026home-mixer/params/param.rs at that commit

The weights

What helps, and what hurts.

In X's published weights, a reply, a quote or a share via DM counts as 7.1 likes.

Helps

  • reply7.1x a like5
  • share via DM7.1x a like5
  • quote7.1x a like5
  • follow the author4
  • sharechanged 5 November 20262.118
  • repost1
  • open a video, when choosing videos for the carouselchanged 15 October 20261
  • likechanged 5 November 20260.708
  • stay 10+ seconds after the tapchanged 5 November 20260.385
  • tap inchanged 5 November 20260.274
  • open link0.2
  • expand a photo0.05
  • pause on the post (dwell)changed 25 August 20260.05
  • tap through to a quoted post0.05
  • unexplored post from an account you followchanged 7 October 20260.015
  • time spent on the postchanged 5 November 20260.0075

Hurts

  • report-234
  • mute the author-58.8
  • not interestedchanged 5 November 2026-53.07
  • block the author-31.2
  • scroll past without pausing-0.02

These are ratios between weights, not engagement counts. A weight multiplies X's predicted probability that a viewer takes the action, so it does not say what one real reply or like is worth.

Guardrails

How to use this when writing.

These are X's published default ranking weights. They describe what the ranker rewards in aggregate. They do not predict what any single post will get.

  1. 01Write the post from the evidence and the creator's voice first. This section never overrides either. If the creator's own best post worked without a device (a question, a hook, a call to action), that is stronger evidence than any weight here.
  2. 02Do not add a question, call to action or hook because a weight rewards it. Replies are weighted highly, but a tacked-on question ("What are you building? Drop it below") reads as a template and risks the costliest signal in the hurts list. A question belongs only if it grows out of the post and is one this audience would genuinely want to answer.
  3. 03Never invent, stretch or generalise a fact to set up engagement. Accuracy comes before every signal.
  4. 04Then run the final check and change the draft only where it fails.

Final check

  • Would the creator actually say this, in these words?
  • Is every claim supported by the brief or the facts provided?
  • Is there a natural reason for someone in this audience to reply, quote it, or send it to one person? The best way is the content itself: a specific number, an honest admission, a finding that surprises. An explicit ask is optional and must pass the rule above.
  • Would anyone outside the audience feel baited or want to hide it? If yes, remove what causes that.
  • If the post is long enough that people tap in to read it, does what follows the tap hold them for 10+ seconds?

For your agent

Give this to your agent.

One line. It reads the markdown version of this page and uses it as a final check on a draft.

Before drafting any X post, read https://www.getlinkintel.com/x-algorithm.md and use it as a final check.

Also as JSON and an RSS feed.

Changelog

Every material weight change.

A weight changes when X edits its code. Each entry links to the commit that did it.

21 material weight changes across 8 commits since 25 August 2026. Last change 13 November 2026.

See the latest change

Further reading

5 November 2026

commit 021afa4 · permalink

14 October 2026

commit 17c96bf · permalink

  • not interested-47.52 → -52.272

    home-mixer/params/param.rs

    The "not interested" penalty got bigger.

    X's published default for the not interested penalty is -52.272, previously -47.52. The penalty got bigger.

  • like0.6 → 0.66

    home-mixer/params/param.rs

    Like counts more.

    X's published default for a like is 0.66, previously 0.6. Likes count more.

7 October 2026

commit 78460ca · permalink

  • unexplored post from an account you follow0.02 → 0.015

    home-mixer/params/param.rs

    Unexplored post from an account you follow counts less: now worth as much as 0.03 of a like, down from 0.04 of a like.

    X's published default for an unexplored post from an account you follow is 0.015, down from 0.02. It counts less, worth as much as 0.03 of a like rather than 0.04.

29 September 2026

commit a707cc2 · permalink

  • not interested-43.2 → -47.52

    vm-ranker/params.rs

    The "not interested" penalty got bigger: now as large as about 95 likes, up from 86.4 likes.

    X's published default for 'not interested' is -47.52, from -43.2. The penalty got bigger, about 95 likes rather than 86.4.

  • stay 10+ seconds after the tap0 → 0.4

    vm-ranker/params.rs

    Stay 10+ seconds after the tap now counts: worth as much as 0.8 of a like.

    X's published default for staying 10+ seconds after a tap is 0.4, from 0. It now counts; a tap followed by that read counts 0.7, from 0.4.

  • tap in0.4 → 0.3

    vm-ranker/params.rs

    Tap in counts less: now worth as much as 0.6 of a like, down from 0.8 of a like.

    X's published default for a tap is 0.3, from 0.4. It counts less than before, worth as much as 0.6 of a like rather than 0.8.

FAQ

Common questions.

What are the X algorithm weights?
X's algorithm weights are the numbers in X's open-source ranking code that score how much each predicted action counts. In the published defaults, a reply, quote or share via DM is weighted 5, a follow the author 4, a repost 1 and a like 0.708. A report is -234. LinkIntel reads these from X's repo every day and links each value to its commit. These are defaults mirrored from X's open-source repo, not proof of what is live, and not a prediction of reach.
Where do the X algorithm weights come from?
They come from X's open-source repo, xai-org/x-algorithm, in the file home-mixer/params/param.rs. The latest synced commit is 590f392, committed 13 November 2026. LinkIntel reads the file every day and links each value to its commit: https://github.com/xai-org/x-algorithm/blob/590f3926e5bc33c50d9ece13953623e1e75ee595/home-mixer/params/param.rs
Do the weights predict how many views a post gets?
No. These are ratios between weights, not engagement counts. A weight multiplies X's predicted probability that a viewer takes the action, so it does not say what one real reply or like is worth. The weights describe what X's ranker rewards in aggregate. Nothing here predicts the reach of a post.
How often do X's ranking weights change?
21 material weight changes across 8 commits since 25 August 2026. The last change was on 13 November 2026. LinkIntel checks the repo every day, and every change in the changelog links to its commit.

Read this before you rely on it

  • The values are defaults mirrored from X's open-source ranking repo. A published default is not proof of what is live for every user.
  • The weights describe what X's ranker rewards in aggregate. Nothing here predicts the reach of a post.
  • LinkIntel is not affiliated with X or endorsed by it.

LinkIntel applies these to your own posts and the reference posts you choose, inside your agent.

Start 7-day free trial