X has taken another major step toward making its recommendation system more transparent, releasing a substantially expanded portion of the code behind its “For You” timeline.
The August 13 update, published on GitHub under the Apache 2.0 license in the xai-org/x-algorithm repository, adds core ranking code, model configurations, filtering systems and content-label mechanisms that influence which posts users see in the platform’s default feed.
The expanded codebase is reportedly 10 to 15 times larger than earlier releases. It offers a much deeper look at how X selects, ranks and filters content, including posts from accounts users follow and recommendations from outside their existing networks.
At the heart of the system are two major content pipelines. In-network posts are handled through X’s Thunder system, while Phoenix and related mechanisms are used to retrieve content from accounts users do not follow.
A transformer model based on Grok then estimates the likelihood that a user will engage with a post. Signals include likes, replies, reposts, quotes, shares, clicks and time spent viewing content. Those predictions are combined with explicit weighting to produce the ranking scores used to determine what appears in the feed.
X Opens the Visibility Filtering Process
The release also provides greater visibility into the systems that can restrict the reach of individual posts.
Depending on the labels assigned to content, the system can allow a post to appear normally, reduce its visibility or place it behind an interstitial warning. The labels can relate to categories including spam, adult material and abusive content, with the underlying models and rule systems now partly available for inspection.
X has also begun testing a new “Under the Hood” transparency tool designed to give some users a clearer explanation of how visibility decisions affect their accounts and posts.
The pilot is currently limited to a randomized group of eligible accounts. Users must have had their accounts for at least a year and posted at least 10 times during the previous month. Those who qualify can inspect labels associated with their accounts or posts, see explanations of what the labels mean and understand how they can affect distribution in the For You feed.
The tool also allows eligible users to download monthly aggregate data, with X expected to broaden access as it collects feedback from the initial test.
Keith Coleman, X’s vice president of product, said the goal is to give the public greater ability to examine and challenge the platform’s recommendation system, including questions about reach, fairness and why particular content is shown.
Elon Musk has similarly framed the broader open-source initiative as a transparency effort. X has previously said it intends to publish its entire codebase after completing security reviews, followed by third-party audits intended to verify that the publicly released code corresponds with the software running in production.
That broader objective had been targeted for August but had not been completed by the middle of the month. The latest algorithm release therefore represents another stage in the process rather than the publication of X’s entire production codebase.
The latest disclosure follows earlier recommendation-system releases in 2026, including an initial release in January and a major update in May. X has invited external experts in recommendation systems to examine the latest code as part of the effort.
Some important elements remain private. Production model weights and certain safety thresholds have not been published, in part because X says exposing them could make the system easier to manipulate.
Even with those limitations, the expanded repository gives researchers and developers a significantly clearer view of how X’s recommendation engine works. Early analysis has focused on the relative importance of positive engagement signals, penalties associated with negative feedback such as reports and mutes, and the mechanics governing the movement of posts through the recommendation pipeline.
The initiative could strengthen X’s position as one of the more transparent major social platforms when it comes to recommendation algorithms. More importantly, it could add pressure on other technology companies to provide greater insight into the systems that determine what billions of users see online.
For X, the challenge will now be ensuring that the published code, the explanations provided to users and the system operating in production continue to match as the platform evolves.
