X released an expanded version of the code behind its recommendation systems on Thursday, including the For You timeline and core ranking engine, and added a settings tool that lets users check which labels have been applied to their account or posts.
X has published this codebase on Github since January 2026 with roughly monthly updates.
The new release adds model configuration, the filtering layer, core ranking details, and the signal weights. The weights were excluded from earlier releases, which xAI attributed to security considerations. Developers reading the repository previously had access to the scoring formula but not the coefficients within it, and published estimates of individual signal values were reverse-engineered rather than sourced from the code.
Stage defaults sit in `home-mixer/params/param.rs`, with scoring logic in the adjacent home-mixer scorer modules. X states the codebase is now roughly ten to fifteen times larger than the previous version; that figure is the company’s own and has not been independently verified.
The release covers the code that retrieves posts, ranks them for a given user, and assembles the feed, along with the systems that filter rule-violating content, and that components including the ranker and scoring logic can be run outside the company.
Ahead of the launch, X gave external recommendation-systems researchers a preview, and they reproduced the per-post score independently.
The account labels tool
A new page in settings, labelled Under the Hood, allows accounts that have posted at least ten times in the past month to download aggregate statistics as a JSON file. The file shows which labels, if any, were applied to the account or its posts over the previous calendar month.
X’s suggested approach for users who do not read JSON is to provide the file and the public repository to an AI assistant and request a plain-language summary.
The tool begins as a pilot restricted to a test group of accounts at least one year old, ahead of any wider rollout.
Exclusions
Systems that use Grok to predict whether a post may violate platform rules are not included in the release. X’s stated reason is that publishing the classifiers would enable evasion.
Those systems cover the categories that generate most moderation disputes. Published code also establishes what a system is designed to do rather than confirming that the code running in production matches the repository, though independent reproduction of the score narrows that gap.
The settings tool reports whether a label was applied to an account. It does not report what the account’s reach would have been without it.
Leave a Reply