How does the TikTok algorithm work in 2026?

TikTok hasn't published its ranking code. It says it weighs your interactions, the video's details and your device and account settings, with strong signals, like watching a longer video to the end, counting more than weak ones. In January 2026 its new U.S. joint venture said it would retrain the algorithm on U.S. user data, kept in Oracle's U.S. cloud.

How does the X algorithm work?

X published its For You code in January 2026. Candidates come from accounts you follow and from a retrieval model; filters drop posts over 48 hours old, ones you've seen and ones from accounts you block or mute; a model predicts your chance of each action; and the score is each chance times a weight, summed, then adjusted for repeated authors and for accounts you don't follow.

Does one report cancel hundreds of likes?

No, though the weights look that way: −234 for a report against 0.66 for a like. X's code points out that the weights multiply your predicted chance of each action, not counts of actions, and that a report's predicted chance is usually more than 1,000 times smaller than a like's. The big weight is what lets that tiny chance count.

Why does my feed keep showing the same kind of post?

What you engage with shapes what's predicted and suggested for you next. In the loop above, liking the cooking posts you're shown turns a mixed feed into all cooking within four sessions. Not interested pushes the other way: on X it carries a weight of −52.272.

What changed from Twitter's 2023 algorithm?

Twitter's 2023 release published its ranking weights: 0.5 for a like, 13.5 for a reply, 75 for a reply the author engages with, −369 for a report. X's 2026 code keeps the weighted sum with new actions and numbers, like 0.66 for a like and 5 for a reply, predicted by a new transformer model.

Live model · The For You feed

How the For You Feed Actually Works

A For You feed gathers candidate posts from accounts you follow and from a model that finds ones you don't, filters out what you've seen or blocked, and predicts your chance of each action on each post: like, reply, share, report. The score is each chance times a weight, summed, and the highest scores fill the feed.

One request, from candidates to feed

One request for your feed. 7 candidates come from accounts you follow and 5 are suggested by retrieval from accounts you don't: 12 in all.

StopFor You ranking

Diversity ruleClip's chance of Not interested

Candidates12Left12Feedtop 5

Candidates

Post	Status	From
Moving day		@maya, following
Miso pasta		@chef_ana, following
Knife skills		@chef_ana, following
Pantry staples		@chef_ana, following
Road closures		@city_news, following
Reunion photos		@old_friend, following
Match tonight?		@you, following
Penalty saves		@viral_clips, suggested
New AI model		@ai_daily, suggested
Late winner		@goal_cam, suggested
90% off		@deals_now, suggested
Miso pasta		@chef_ana, suggested

X's weights, one per predicted action

Action	Chance	Weight	Adds
Like		0.66
Reply		5
Repost		1
Share		2
Follow author		4
Not interested		−52.272
Report		−234
Score

Scores

(none)

Watch the Python behind this model run, line by line

How to read the stage

Twelve candidates arrive for one request: seven from accounts you follow, five suggested. Filters drop a duplicate, a 72-hour-old post, your own, one you've seen and one from a blocked account. Each of the seven left is scored with X's published weights, then multiplied down for repeat authors and, by 0.75, for accounts you don't follow; the top 5 make the feed. Maya's moving-day post wins at 0.577, mostly on its 10% chance of a reply, worth 0.500.

Why a likely like isn't enough

The viral clip has the best chance of a like of any post, 30%, which adds 0.198. But its 2% chance of Not interested costs 2% × −52.272 = −1.045, five times as much, and the clip finishes last at −0.624. X weights a reply at 5 and a report at −234. Its code explains why rare actions get big weights: a report is so unlikely that a small weight couldn't move the ranking at all.

Why one account can't fill your feed

Chef Ana has three posts in the running. Her first counts in full; X multiplies each later one by 0.75 × 0.5^k + 0.25, so her second keeps 0.625 of its score and her third 0.4375. That trades her pantry post for an AI post. Turn the rule off under Break it and three of the five posts are hers. TikTok says its feed generally won't show two videos in a row from the same creator.

The feedback loop

X's predictions for you are driven largely by your own behavior. In the stage below, a reader likes every cooking post shown, and in this toy each like adds 0.15 to their interest in cooking, raising those posts' chances and their share of suggestions. Session 1 shows 1 cooking post of 4; by session 4, all 4 are cooking, though the weights never change. Set the boost to 0 and the mix stays.

Six sessions of one reader

Six sessions of one reader. Each sources 10 candidates: one per topic from accounts you follow and 6 suggested, shared out by your interests. All score with X's weights; the top 4 make the feed. You like every cooking post you see; each like: +0.15.

StopThe feedback loop

SeedInterest per like

Session0 of 6Cooking interest1.00×

Your feed

#	Post	Topic	Score	You
(none)

Candidates by topic: following, then suggested

Posts in your feed, by topic

Cooking
Soccer
AI
News

Your predicted interest

CH1 Cooking
CH2 Soccer
CH3 AI
CH4 News

Watch the Python behind this model run, line by line

What the model leaves out

Every chance on this page is stated. X predicts over 20 actions per post with a transformer trained on engagement, and finds suggested posts with machine-learned retrieval; the toy's interest rule stands in for both. It also skips X's other filters, its new-author boost, a reranking step and a small score offset. TikTok's ranking code isn't public, so this page models X's.

Questions

How does the TikTok algorithm work in 2026?: TikTok hasn't published its ranking code. It says it weighs your interactions, the video's details and your device and account settings, with strong signals, like watching a longer video to the end, counting more than weak ones. In January 2026 its new U.S. joint venture said it would retrain the algorithm on U.S. user data, kept in Oracle's U.S. cloud.
How does the X algorithm work?: X published its For You code in January 2026. Candidates come from accounts you follow and from a retrieval model; filters drop posts over 48 hours old, ones you've seen and ones from accounts you block or mute; a model predicts your chance of each action; and the score is each chance times a weight, summed, then adjusted for repeated authors and for accounts you don't follow.
Does one report cancel hundreds of likes?: No, though the weights look that way: −234 for a report against 0.66 for a like. X's code points out that the weights multiply your predicted chance of each action, not counts of actions, and that a report's predicted chance is usually more than 1,000 times smaller than a like's. The big weight is what lets that tiny chance count.
Why does my feed keep showing the same kind of post?: What you engage with shapes what's predicted and suggested for you next. In the loop above, liking the cooking posts you're shown turns a mixed feed into all cooking within four sessions. Not interested pushes the other way: on X it carries a weight of −52.272.
What changed from Twitter's 2023 algorithm?: Twitter's 2023 release published its ranking weights: 0.5 for a like, 13.5 for a reply, 75 for a reply the author engages with, −369 for a report. X's 2026 code keeps the weighted sum with new actions and numbers, like 0.66 for a like and 5 for a reply, predicted by a new transformer model.

Drill the mechanism

Filling the feed is top-k selection: keep the 5 highest of many scores. X's selector sorts every candidate and takes the first ones; a min-heap of size k does it in one pass without sorting the rest, which is what Kth Largest Element in an Array and Top K Frequent Elements drill.