X: the only platform that publishes its ranking code, and what the code says

Two open repositories, two different systems.

X is the only platform whose ranking code you can read. What is less understood is that there are two versions of it in public, they disagree, and the older one is still presented as current.

What the release openly is not

"This is a mini version of the Phoenix model (128-dim, 4-layer transformer)... Production uses a larger model with more layers and wider embeddings." And: "This release is a frozen checkpoint from that continuous training process, a snapshot at a point in time."

1. GitHub, twitter/the-algorithm , README and commit history. Verified by clone, 8 August 2026. 1st
2. GitHub, xai-org/x-algorithm , README and phoenix/README.md , 15 May 2026. 1st
3. Musk quoted by Reuters, 10 January 2025, and by BeInCrypto, 15 May 2026. 2nd

The weights everyone quotes, and their date.

These ten numbers appear in almost every article about X. They are the only numeric ranking weights X has ever published, and they were published on 5 April 2023.

X's formula, verbatim: "score = sum_i { (weight of engagement i) * (probability of engagement i) }." ¹¹

What happened to them since

In the September 2025 release every engagement weight default was set to 0.0 . Twenty-nine parameters, all zeroed. No weights are disclosed in that version. ¹²

In the May 2026 release the weights are referenced by name throughout the scoring code and the module that defines them is not in the repository. A search for FAVORITE_WEIGHT returns exactly one file, the scorer that consumes it. The release does not compile as published. ¹³

1. GitHub, twitter/the-algorithm-ml , Heavy Ranker README, commit dated 5 April 2023. 1st
2. GitHub, twitter/the-algorithm , HomeGlobalParams.scala , 3 September 2025. Verified by clone. 1st
3. GitHub, xai-org/x-algorithm , weighted_scorer.rs and lib.rs , 15 May 2026. Absence of the params module verified by clone and repository-wide search. 1st

What the current system does, where it is readable.

The numbers are missing but the mechanisms are not. Several of them are published nowhere else by any platform in this document.

The fifteen predicted actions

Favorite, reply, repost, quote, click, profile click, video view, photo expand, share, dwell, follow author, not interested, block author, mute author, report. "Positive actions have positive weights. Negative actions have negative weights, pushing down content the user would likely dislike." ¹⁴

The scorer in the code actually sums nineteen terms, more than the README lists. The extras are worth naming: separate weights for a share by direct message, a share by copied link, a click on a quoted post, and continuous dwell time. ¹⁴

Video has a minimum duration gate

Video engagement only contributes to the score if the video exceeds a minimum duration. The threshold is one of the unpublished values. ¹⁴

Two mechanisms nobody else publishes

Author diversity decay. Within a single feed response, each additional post from the same author is multiplied down by a decay factor. The code comment is plain: "Diversify authors served within a single feed response." Both the decay and the floor are unpublished. ¹⁴

An out of network penalty. Posts from accounts the viewer does not follow are multiplied by a penalty factor after scoring, with the comment "Prioritize in-network candidates over out-of-network candidates." In the September 2025 Scala code that factor was published at 0.75, alongside a reply penalty also at 0.75. In the 2026 code the factor exists and the number does not. ^{14 , 15}

1. GitHub, xai-org/x-algorithm , README, weighted_scorer.rs , author_diversity_scorer.rs , oon_scorer.rs , ads/util.rs . 15 May 2026. 1st
2. GitHub, twitter/the-algorithm , ScoredTweetsParam.scala , 3 September 2025. 1st

Links, and the owner contradicting himself.

X the company has published nothing on this. Everything attributed to X comes from Musk's personal posts, and two of them do not reconcile.

October 2023

"Our algorithm tries to optimize time spent on X, so links don't get as much attention, because there is less time spent if people click away."

November 2024

"Just write a description in the main post and put the link in the reply. This just stops lazy linking."

April 2025

"To be clear, there is no explicit rule limiting the reach of links in posts. The algorithm tries (not always successfully) to maximize user-seconds on X, so a link that causes people to cut short their time here will naturally get less exposure."

What the code says

Neither release contains a link based ranking penalty. In the 2026 code the only link related term anywhere in the scoring path is a share by copied link, which is a positive action. There is no has-link feature, no URL penalty, no external domain feature and no link filter. ¹⁹

The one thing that did exist: the 2025 code carried a filter that removes out of network posts containing URLs on a configurable competitor list. The list ships empty. Whatever is in it in production is not disclosed. It is domain specific and out of network only, and it is not in the 2026 code at all. ²⁰

1. Axios, quoting Musk, 3 October 2023. 2nd
2. Mediaite, quoting Musk, 25 November 2024. 2nd
3. Musk on X, 25 April 2025. x.com blocks automated retrieval, so this was read from a search index rather than the live post. 1st
4. GitHub, xai-org/x-algorithm , full repository search, 15 May 2026. 1st
5. GitHub, twitter/the-algorithm , OutOfNetworkCompetitorURLFilter.scala , 3 September 2025. 1st

Premium, and what Grok now decides.

One of these is smaller than people think. The other is larger, and almost nobody is talking about it.

What Premium claims, in X's own words

The help page claims exactly one ranking benefit: "people on X will see a slight preference for replies from verified accounts over other replies." It adds that "we're currently testing the levels at which we prioritize content from Premium subscribers relative to the other factors we consider in conversation rankings." It claims no For You timeline reach benefit at all. The page carries no date. ²¹

Grok is now inside the ranker

The 2026 model's transformer is "ported from the Grok-1 open source release by xAI, adapted for recommendation system use cases." ²³

The May 2026 release added Grox, a content understanding service running classifiers for "spam detection, post-category classification, and PTOS policy enforcement." ²³ One of those classifiers is named for the accounts it is applied to: a spam classifier specifically for low follower users. The prompt defining what it treats as spam is not published. ²³

The 2025 code went further and is worth knowing about: Grok annotations hard filtered posts for spam, gore, NSFW and violence, and a separate rescorer applied a decay to content Grok labelled as slop. The decay ships at 1.0, meaning no effect, and the production value is somewhere between 0 and 1 and is not disclosed. Grok also assigned per-post features including a political inclination score. ²⁴

1. X Help Centre, About X Premium . No date on the page. Retrieved 8 August 2026. 1st
2. GitHub, twitter/the-algorithm , ScoredTweetsParam.scala at the 2023 state and at 3 September 2025. Verified by clone. 1st
3. GitHub, xai-org/x-algorithm , README and grox/classifiers/content/ , 15 May 2026. 1st
4. GitHub, twitter/the-algorithm , GrokSlopScoreRescorer.scala , GrokSpamFilter.scala , HomeFeatures.scala , 3 September 2025. 1st

What a view is, and what the numbers are worth.

X's view count is not reach. Its own two definitions of the same event do not match. And it publishes no audited user figure at all.

X's definition of a view

"Anyone who is logged into X who views a post counts as a view, regardless of where they see the post or whether or not they follow the author. Multiple views may be counted if you view a post more than once, but not all views are unique."

No dwell requirement is stated. A scroll past registers. Logged out viewers are excluded. It is closer to gross impressions including repeats than to reach.

User numbers

Musk stated in February 2026 that "our monthly users are, on average, around 600 million," and that "the number of people who have the X app installed is well over a billion." ²⁷ X has been private since October 2022 and files no quarterly reports. There is no published methodology, no definition of a monthly user and no audit.

The one auditable X figure is its average monthly active recipients in the European Union, published under Digital Services Act obligations. It covers the EU only and does not substitute for a UK or US number. ²⁸

Measured engagement, 2025 full year

2,711 impressions per post and 42.71 interactions per post, giving an engagement rate of 1.58 percent per impression, up from 1.32 percent, while impressions fell 5 percent. 1,123,528 posts from 15,116 accounts. ²⁹ As with every panel in this document, these are the customers of a scheduling tool, not a sample of X.

1. X Help Centre, View counts . No date. 1st
2. X Business, Post and Video Activity Dashboards . No date. 1st
3. Musk, quoted verbatim by Social Media Today, 12 February 2026. 2nd
4. X Transparency Centre, AMARS in the EU . Page exists; contents not retrieved in this research. 1st
5. Metricool, X Stats 2026 , 26 February 2026. Denominator per impression, confirmed by arithmetic. 2nd