Inside X’s Open‑Source Ranking Algorithm: Boosting Transparency and Code Quality

Open‑source release

The platform released the full source code of its personalized recommendation and ranking engine on GitHub (https://github.com/xai-org/x-algorithm) under the Apache v2 license. The codebase contains roughly 300,000 lines and replaces the 2023 Scala implementation with a Rust and Python rewrite.

A commit dated 15 May 2026 modified more than 18,000 lines across 187 files, expanding the repository size by about 10–15 times.

Ranking pipeline

The engine operates in multiple stages. First, it gathers candidate posts from three sources: accounts the user follows, trending topics, and content similar to the user’s past interactions. The ranking model then predicts the probability of user interaction for each candidate and sorts the candidates accordingly.

Posts that receive shares, comments, replies, or other positive engagement signals are promoted, while posts that are reported, muted, or receive little attention are penalised.

Model architecture

The core ranking model is a Transformer named Phoenix , which incorporates elements from the platform’s Grok technology. The pipeline also includes visibility filters, safety modules, and classification layers that process content‑moderation signals.

Some components, such as the Grok‑based rule‑violation predictor, remain closed‑source to prevent malicious actors from bypassing platform rules.

Transparency tools

A new “Under the Hood” tool exposes configuration parameters and provides tag‑level statistics. Users with at least ten posts in the past month can download a JSON file summarising whether and how their posts were affected by ranking tags.

The tool is initially offered as a pilot to accounts registered for at least one year before broader rollout.

Community involvement

External researchers received a preview of the open‑source code, allowing them to train and run the Phoenix scoring system. Developers can submit pull requests to the GitHub repository; engineers may incorporate accepted contributions into the production algorithm.

Context among platforms

By publishing a runnable recommendation engine, the platform joins a small group of mainstream social networks that provide full source code. Other major platforms (Meta, TikTok, YouTube) have only released research papers or algorithm sketches, not complete codebases.