How the X (Twitter) algorithm works in 2026: the real weights

X publishes its For You ranking code and keeps the weights in sync with production. We read it on 20 October 2026: what each action is worth, what changed since August, and why replies from strangers never reach For You.

X’s For You algorithm predicts how likely you are to take each of about thirty actions on a post (reply, copy its link, follow the author, mute, report) and ranks posts by a weighted sum of those predictions. The weights are public. X publishes its ranking code, and the README says the values in it are kept in sync with production, so you can read them for yourself.

We read the repository on 20 October 2026, at commit 74d49a4. Copying a post’s link is worth the most, a reply is worth about 7.6 likes (the 27x you’ll find elsewhere is from 2023), and replies from accounts you don’t follow are filtered out of For You altogether. Several weights have changed since August, and they are listed by date below.

How does the X algorithm work in 2026?

Each time you open For You, X builds the feed in five steps. The names are the ones in the code.

1

Collect candidates — Recent posts from accounts you follow, plus posts from accounts you don’t follow that a retrieval model picks for you.

2

Filter — Out go posts older than 48 hours, muted words and accounts, posts you’ve seen, and reposts and replies from accounts you don’t follow.

3

Predict — A model X calls Phoenix, built on the same transformer design as xAI’s Grok, predicts the probability of each action: like, reply, repost, quote, copy link, follow, mute, report and the rest.

4

Score — Each probability is multiplied by its weight and summed. Posts from accounts you don’t follow are then multiplied by 0.75, and each extra post from the same author in one feed is scored lower.

5

Order for variety — A final pass reorders the top posts so the feed isn’t ten near-identical ones in a row.

The README puts the score as Final Score = Σ (weight × P(action)). So a post doesn’t gain from likes it already has. It ranks well for you when the model expects you to do something valuable with it.

The weights are the live ones, with one caveat. X first published the new algorithm on 20 January 2026 and added the ranking weights, filters and content labels on 13 August (@XOpenSource). The README says the numbers are the ones production runs:

The README of xai-org/x-algorithm, section “Experiments and Configuration”: experiments run on a small percentage of traffic, the aim is for those at 10% or more to be visible in the repository, and “we run cron scripts that set the defaults in this repository’s code to be the primary production values, for example in home-mixer/params/param.rs.”
The README on what the numbers in the code are. xai-org/x-algorithm, README, 20 October 2026.

The caveat is in the same paragraph: experiments on less than about 10% of traffic may not show up. The trained model itself isn’t published either, so the code tells you what X rewards, but not how the model judges a given post. The weights file changed in 44 commits between 13 August and 20 October 2026 (history), so any number here, ours included, has a date on it.

What does the X algorithm reward most?

Copying a post’s link, by a wide margin. Then replies, quotes and sending the post in a DM, then following the author. A like is one of the cheapest signals:

For You weights by action, xai-org/x-algorithm, 20 October 2026
Copy link20
Reply on a mutual’s original post20 (5 + 15)
Reply5
Quote5
Share via DM5
Follow the author4
Repost1
Like0.66
Dwell (stopped on the post)0.05
Profile click0
Lines 374 to 388 of home-mixer/params/param.rs on GitHub: ShareWeight 2.0, ShareViaDmWeight 5.0, ShareViaCopyLinkWeight 20.0, DwellWeight 0.05, QuoteWeight 5.0.
Where the copy-link weight lives: param.rs, lines 374–388. param.rs at 74d49a4, 20 October 2026.

Read the ratios with care. The README warns that weights multiply the predicted probability of an action, not counts, so “one report cancels out 468 likes” is a misreading. The chart still shows where the weight sits: sending a post on, through a copied link (20) or a DM (5), counts for far more than liking it (0.66). The extra 15 on a reply applies only to original posts (not replies or reposts) by an account the viewer follows back. Nikita Bier announced it in July as a boost for your mutuals.

What changed in the X algorithm since August 2026?

Most guides still quote the numbers from August, or from 2023. These are the changes to the weights and rules since X published them on 13 August, by the date of the commit:

DateWhat changed
25 AugustDwell came back: 0 → 0.05. Video quality views dropped to 0.
24 SeptemberThe old ranking scorer was deleted. The 0.75 for accounts you don’t follow and the author-diversity penalty moved to a separate service.
29 SeptemberTime spent on a post after opening it now counts (0 → 0.4); a click dropped from 0.4 to 0.3.
30 SeptemberCold start rebuilt: authors up to 50,000 followers instead of 1,000, posts under 200 views and at most 2 hours old.
13–14 OctoberLike 0.5 → 0.6 → 0.66. “Not interested” went from −43.2 (August) to −52.27.

What didn’t move: copy link 20, reply 5, the mutuals boost 15, quote and DM 5, follow 4, repost 1, report −234, mute −58.8, block −31.2, and the 0.75 for accounts you don’t follow. On 3 October X also added a switch that would let replies from accounts you don’t follow into For You. It is off.

Do replies help you grow on X?

Not through For You. A filter drops every reply from an account the viewer doesn’t follow before scoring, and the switch that would let some through is set to false:

Lines 9 to 32 of home-mixer/filters/oon_retweet_reply_filter.rs on GitHub: the filter reads EnablePhoenixOonReplies and PhoenixOonRepliesRootOnly, marks a candidate as a reply or a repost, and removes it when in_network is false and it is a repost or a dropped reply.
The reply filter. With EnablePhoenixOonReplies false, a reply from an account the viewer doesn’t follow is removed. oon_retweet_reply_filter.rs at 74d49a4, 20 October 2026.

Even your followers see your replies at 0.75 of their score (vm-ranker params), and neither the mutuals boost nor cold start applies to a reply. So a reply reaches strangers one way only: under the post you replied to, where everyone reading that thread scrolls past it. X’s help page on replies says it ranks a reply higher when the post’s author replied to it, when the viewer follows you, or when you have Premium.

What a good reply under a post your market reads does get you is profile visits and follows, and a follow is what puts your next original post into that person’s For You. How to pick the posts and write the replies is in the reply guy strategy.

Replies under the posts your market reads, and posts that earn the follow, from your own account.

Book a call

Can a small account still get reach on X?

Yes, and since 30 September far more accounts qualify. X calls it cold start. An original post (not a reply or a repost) from an author with up to 50,000 followers, under 200 views in For You and no more than two hours old can be lifted to the score of the 15th or 16th post in a stranger’s feed:

Lines 531 to 560 of home-mixer/params/param.rs on GitHub: ColdStartImpressionThreshold 200, ColdStartSlotMin 15, ColdStartSlotMax 16, ColdStartFollowerCap 50000, ColdStartMaxPostAgeSecs 7200.
Cold start: 200 views, slot 15–16, 50,000 followers, 7,200 seconds. param.rs at 74d49a4, 20 October 2026.

Only one such post is lifted per feed, and X picks it by sampling on the share of viewers who liked it so far, so a post that gets early likes is more likely to be tried again. The 0.75 for strangers still applies on top. In August the same rule covered only accounts under 1,000 followers and posts up to 24 hours old. For an account under 50,000 followers, the first two hours are the window for this lift.

What hurts reach on X?

The negative actions, weighted far above any positive one. As with the positive weights, they multiply each viewer’s predicted chance of doing it:

Negative For You weights, shown without the minus sign, 20 October 2026
Report−234
Mute the author−58.8
“Not interested”−52.27
Block the author−31.2

Posting a lot at once has a cost too: within one feed, each further post by the same author is scored at 0.625, then 0.44, down to a floor of 0.25 of its score. Content labels can keep posts out of recommendations entirely. How to see yours, and what to do about them, is in our X shadowban check. On links: the ranking code has no rule against them, and opening a link is a small positive weight (0.2). The model may still have learned to rank them lower, and a study of more than 40 million posts (NDSS 2026) found links cut visibility by up to a factor of eight.

Which X algorithm “facts” are myths in 2026?

These come up in guides that rank for “twitter algorithm” today. Most come from the code Twitter published in 2023, which the new system replaced:

What you’ll readWhat the code says on 20 October 2026
“A reply is worth 27x a like”Reply 5.0, like 0.66: about 7.6x. 30x only on a mutual’s original post.
“An author reply is worth 75x or 150x”There is no such action in the scoring.
“The first 30 minutes decide everything”No such window. Cold start covers posts up to 2 hours old.
“Premium gives a 2–4x boost”No Premium multiplier in For You scoring. X’s help pages say Premium ranks replies higher.
“TweepCred scores your account”TweepCred is not in the 2026 code.
“Links cut reach by 50%”No rule in the code. The model may have learned it.

Check the date on any guide before acting on it, this one included. If it quotes a weight without a date or a commit, it’s probably repeating someone else’s.

Frequently asked questions

Is the X algorithm open source?

Yes. X published its new For You algorithm on GitHub as xai-org/x-algorithm in January 2026 and added the ranking weights, filters and content labels on 13 August. The README says cron scripts keep the weights in the code equal to the production defaults. The trained model and experiments on under about 10% of traffic are not published.

What does the X algorithm value most?

On 20 October 2026, copying a post’s link carried the highest weight, 20. A reply, a quote and sharing in a DM were 5 each, following the author 4, a repost 1 and a like 0.66. A reply on an original post by someone the viewer follows back gets an extra 15. Weights multiply predicted probabilities, not counts.

Does the X algorithm penalize links?

Not by a rule. The ranking code in xai-org/x-algorithm has no link penalty, and opening a link carries a small positive weight, 0.2. The prediction model may still have learned that posts with links get fewer replies and likes, and an NDSS 2026 study of more than 40 million posts measured up to eight times less visibility.

Does X Premium boost your reach in the algorithm?

Not in For You scoring. The 2026 code has no Premium or verified multiplier there, although verification is one of the inputs to the model that ranks replies. X’s help page on replies says Premium accounts rank higher in reply threads, so Premium helps where your replies are read, not in strangers’ For You feeds.

How often does the X algorithm change?

The weights file in xai-org/x-algorithm changed in 44 commits between 13 August and 20 October 2026, mostly by an automated sync. The like weight moved from 0.5 to 0.66, dwell came back, and on 30 September cold start was extended to authors with up to 50,000 followers. Check the date on any weight you read.

Replies that earn the follow, posts that earn the reach.

Daily replies from your own account, in your voice, under posts your market already reads.

Run adds a content plan and up to 14 posts a week, the original posts For You can lift.

You approve examples of your voice and the accounts to reply under before the start.

Access through X’s delegate settings, no password. Remove us on the same screen at any moment.

Work with usDone-for-you X service