Inside Twitter’s Open‑Source Recommendation Algorithm: How Likes, Retweets, and Follows Are Weighted

Elon Musk fulfilled his promise by open‑sourcing Twitter’s recommendation algorithm on GitHub, quickly gathering nearly 30,000 stars.

Before the release, the tech community was buzzing, and Sourcegraph even hosted a live stream.

Below is a quick walkthrough of the core ranking weight rules extracted from the source file

https://github.com/twitter/the-algorithm/blob/main/cr-mixer/server/src/main/scala/com/twitter/cr_mixer/similarity_engine/EarlybirdTensorflowBasedSimilarityEngine.scala#L142

:

retweetCountParams – retweets count 20× weight.

replyCountParams – replies count 1× weight (20 replies ≈ 1 retweet).

reputationParams – reputation 0.2× weight.

luceneScoreParams – Lucene scoring 2× weight (content relevance).

textScoreParams – pure text 0.18× weight.

urlParams – URLs 2× weight.

isReplyParams – replies to others 1× weight.

favCountParams – likes 30× weight.

langEnglishUIBoost – English UI 0.5× weight.

langEnglishTweetBoost – English tweets 0.2× weight.

langDefaultBoost – default language 0.02× weight.

unknownLanguageBoost – unknown language 0.05× weight.

offensiveBoost – offensive content 0.1× weight (penalized).

inTrustedCircleBoost – trusted‑circle content 2× weight.

multipleHashtagsOrTrendsBoost – multiple tags 0.6× weight.

inDirectFollowBoost – indirect follow (author follows after seeing tweet) 4× weight.

tweetHasTrendBoost – trending topics 1.1× weight.

selfTweetBoost – self‑retweet 2× weight.

tweetHasImageUrlBoost – image URL 2× weight.

tweetHasVideoUrlBoost – video URL 2× weight.

In short, Twitter’s “one‑click three‑likes” are likes (30×), retweets (20×), and follows (4×).