Does X still use engagement weights like "replies = 13.5x"?
No. The famous numbers — replies worth 13.5x a like, a report worth −369 — come from the 2023 open-source release, and the system that used them has been replaced. The current open-source X algorithm (released January 2026, updated May 2026) still combines predicted engagements with a weighted sum, but the 2023 values are obsolete and the current numeric weights are not published in the code.
Where the 13.5x figure came from
In March 2023, Twitter open-sourced a snapshot of its recommendation system. The companion ML repository documented the "heavy ranker" — the model that scored candidate tweets — and published the exact weights applied to each predicted engagement, dated April 5, 2023 in the source itself:
In the 2023 open-source release, the heavy ranker's published weights (dated April 5, 2023 in the source) were: favorite 0.5, retweet 1.0, reply 13.5, good profile click 12.0, video playback (50%) 0.005, reply engaged by author 75.0, good click 11.0, good click v2 10.0, negative feedback v2 −74.0, report −369.0.
This table is the origin of nearly every "the algorithm rewards X" claim you have read since. It was real, it was citable, and for a while it was a genuinely useful map. The structural idea was a weighted sum: predict the probability of each engagement, multiply by its weight, add it up.
The 2023 heavy ranker also combined predicted engagement probabilities into a weighted sum — the same structural idea the 2026 system retains, with different models producing the probabilities and different (unpublished) weights.
What replaced it
In January 2026, X Engineering open-sourced its new algorithm — described by X as the live production system, built on the same transformer architecture as xAI's Grok model, with public updates committed roughly every four weeks.
In January 2026, X Engineering open-sourced its production For You feed algorithm at xai-org/x-algorithm, built on the same transformer architecture as xAI's Grok model.
The new system's own documentation is unambiguous about the break with the 2023 design: the hand-engineered features are gone. A Grok-based transformer called Phoenix learns what is relevant to you from your engagement history, rather than from manually constructed feature pipelines.
The current X algorithm has eliminated every hand-engineered feature and most heuristics; the Grok-based transformer learns relevance from the user's engagement history.
What the current code actually shows
Here is the nuance most coverage gets wrong in both directions. The weighted-sum structure did not die — the current code still combines predicted engagement probabilities into a final score.
Phoenix, a Grok-based transformer, predicts per-post probabilities for many engagement types; the final ranking score is a weighted combination of those predicted engagements.
The current scorer weighs a richer set of predicted actions than 2023 did — including dwell time, photo expands, shares via DM, and shares via copied link:
The current WeightedScorer combines predicted probabilities for: favorite, reply, retweet, photo expand, click, profile click, video quality view, share, share via DM, share via copied link, dwell, quote, quoted click, continuous dwell time, follow author, not interested, block author, mute author, and report.
And the negative side is explicit in the code: the model predicts whether you will hit "not interested," block the author, mute the author, or report the post — and those predictions carry negative weights that push content down.
The model explicitly predicts negative actions — not interested, block author, mute author, report — and these carry negative weights in the final score, pushing down content a user would likely dislike.
What we can't verify
The one thing the current release does not contain is the numbers. The scorer references its weights from a parameters module — p::FAVORITE_WEIGHT, p::REPLY_WEIGHT, and so on — and that module's values are absent from the published repository. We checked the full tree at the pinned commit. Anyone quoting current numeric weights for the X algorithm is guessing.
The numeric values of the current weights are not included in the open-source release: weighted_scorer.rs references a params module (e.g. p::FAVORITE_WEIGHT, p::REPLY_WEIGHT) whose values are not present anywhere in the published repository.
It is also worth scoping honestly: "the open-source repo is the production system" is X's own characterization. The code is the best public evidence that exists, but only X can prove the deployed system matches it at any given moment.
What this means for you
Stop optimizing for the 2023 table — anyone selling you a "13.5x replies" strategy is working from a superseded system. What the current code does give you is a precise map of which actions matter: replies, reposts, quotes, profile clicks, dwell, shares (including private shares via DM and copied links), and follows on the positive side; "not interested," blocks, mutes, and reports on the negative side. The practical strategy that survives the rewrite is simple to state: maximize the probability of the positive actions and minimize the probability of the negative ones — because that is, literally and citably, the equation.