X dropped the code behind its For You feed into the open on Thursday, and it shipped with a tool that finally answers the question users have been asking for years: is the X algorithm quietly throttling your posts? For once, the answer isn't a rumor, it's a repository you can read.
The For You feed code now lives on GitHub under an Apache 2.0 license, and it includes the weights that help decide which posts you see. The public repo has also grown substantially since its January debut: a May update added a runnable end-to-end inference pipeline, additional content-understanding components, ads blending, and more candidate sources.
I covered the original January open-source announcement when X first laid out the plan. What is public now gives us a much clearer view of how the major pieces fit together.
There's also a new transparency tool called Under the Hood that lets participating users download aggregate information showing whether labels have affected their visibility.
What the X algorithm actually released
The release covers the For You timeline end to end: model configuration, the core ranking system, filters, and parameters that weight different predicted signals. The repo lives under the xai-org organization on GitHub, reflecting X's 2025 combination with xAI. xAI, including X, was later acquired by SpaceX in February 2026.
The architecture is unusually readable for a production-scale social feed. Home Mixer orchestrates the timeline, Thunder supplies recent posts from accounts you follow, and Phoenix handles out-of-network retrieval and ranking with a Grok-based transformer model.
The repo has also been updated since its January launch. The May 15 release added a single retrieval-to-ranking pipeline, a packaged mini Phoenix model, content-understanding services, ads blending, and additional candidate sources. Open source doesn't mean every production detail is exposed, but there is enough here to inspect how the major pieces fit together.
That also puts the release in context with Musk's July promise to open source X's entire codebase. The recommendation stack is now far more inspectable than it was at the start of the year, even if "everything" is still a bigger claim than what is publicly exposed today.
What the X algorithm actually does
Ranking and visibility are separate parts of the system, and that split is the key to understanding reach. The X algorithm ranks posts in order, while filtering layers decide whether particular candidates remain eligible to be shown.
That's how a post can rank well in one context and still lose distribution because a separate eligibility or visibility rule applies.
The published weights show how the X algorithm scales predicted actions against each other. A predicted report, for instance, carries a negative weight about 468 times the magnitude of a predicted like. Mutes, blocks, and "not interested" predictions push the score the other way too.
That does not mean one report literally cancels out 468 likes. These weights multiply predicted probabilities produced by the model; they aren't a simple scoreboard of raw engagement counts.
The visibility filtering system also applies rules that can remove posts from recommendation surfaces for reasons such as spam, violence, gore, blocked or muted relationships, previously seen content, and other eligibility conditions. That's one of the mechanisms behind what users often describe as being "shadowbanned," although the public code doesn't reduce that idea to one universal on/off switch.
There is an important caveat here. The public Phoenix documentation explicitly describes the released model as a smaller, frozen snapshot, while production uses a larger model trained continuously on real-time data. So this is real source code and a useful representation of the system, but it is not a perfect copy of every production model weight and runtime detail.
How to check the X algorithm for visibility limits
The practical part is Under the Hood, a page X is piloting with a randomized test group of eligible accounts. Participating users can see aggregate information about labels applied to their account and posts that can affect visibility.
That means you may not have access yet. X says a wider rollout will depend on feedback from the pilot.
X's suggested workflow is the most 2026 thing imaginable: take the data, compare it with the public algorithm code, and use an LLM to help explain what the labels and ranking logic mean. You can check the Under the Hood page to see whether the tool is available on your account.
The honest caveat is that this is aggregate transparency, not a forensic per-post audit. It can tell you far more than X exposed before, but it still doesn't turn every mysterious impression drop into a neatly labeled cause.
X algorithm caveats
This is a big step for a platform that has historically given users very little visibility into recommendation mechanics, but it's not the whole story. A published repository is not proof that every line, model weight, configuration value, and production service matches what is running live at any given moment.
The political history matters here too. Claims about hidden reach suppression have been part of the Twitter and X argument for years, especially around politically sensitive accounts and content.
Now users and researchers have much more code to inspect instead of relying entirely on platform statements or outside speculation. That's a real change, even if open source by itself can't settle every argument about how the live service behaves.
The other tension is that X can be more open about recommendation code without becoming an open company in every other sense. X became part of xAI in 2025, and xAI Holdings was acquired by SpaceX in February 2026. Corporate consolidation and source-code transparency are two different things.
What to watch
Watch the Under the Hood pilot expand beyond the test group, and watch whether X keeps publishing meaningful updates to the repository over time.
Also watch the gap between "open source" and "open to contribution." The repository is public and can be read, cloned, forked, and audited, but GitHub currently shows new pull request creation as restricted on the xai-org/x-algorithm repo.
For anyone who has ever suspected their reach was being throttled, the short version is this: there are finally more receipts in public than there used to be, and X is giving at least some users a way to compare those receipts with what happened to their own account.
