𝕏 Just Open-Sourced Their Algorithm Again: What Changed and What It Means

𝕏 dropped their algorithm source code this morning. Not a blog post summarising it, not a press release with selected highlights, but the actual code that decides what 500 million daily users see in their feeds.

The last time they did this was March 2023. Back then, we learned specific numbers: replies weighted at 13.5x compared to likes, author replies triggering a 75x multiplier, retweets worth 20x. Those numbers shaped every serious growth strategy for the past three years.

The 2023 system is gone. They've rebuilt the algorithm from scratch.

The Architecture Has Changed Fundamentally

The 2023 algorithm ran on a ~48 million parameter neural network with hand-engineered features. Engineers manually decided which signals mattered and assigned specific weights to each one. SimClusters grouped users into 145,000 communities based on engagement patterns. TweepCred scored accounts from 0-100 based on reputation signals.

None of that infrastructure appears in the new codebase.

The new system runs on Phoenix, a Grok-based transformer model adapted from xAI's open source release. It's the same architecture powering Grok AI, repurposed to decide what shows up in your feed.

The most significant change is philosophical as much as technical: 𝕏 has eliminated every hand-engineered feature. No more engineers deciding that a reply should be worth exactly 13.5x a like. The transformer learns relevance directly from engagement patterns. It watches what you like, reply to, share, and scroll past, then predicts what you'll engage with next based on those patterns rather than predetermined rules.

How the New System Works

The algorithm now operates in two distinct stages.

The first stage is retrieval. A Two-Tower model creates embeddings for users and posts. Your embedding encodes your engagement history, who you follow, and your demonstrated preferences. Every candidate post gets its own embedding based on content and author. The system finds posts most similar to your interests using approximate nearest neighbor search, narrowing millions of candidates down to roughly a thousand.

The second stage is ranking. The Grok-based transformer takes those thousand candidates and predicts the probability you'll take specific actions on each one. Not a single "will they engage" score, but separate predictions for each action type: will you like it, reply to it, repost it, click it, share it via DM, expand the photo, watch the video, follow the author, or hit "not interested."

Each action type has a weight. Positive actions add to the final score. Negative actions subtract from it. The weighted sum determines where each post appears in your feed.

The Weights Are No Longer Public

In 2023, we could read the exact multipliers in the source code. A reply was worth 13.5 points. An author reply triggered a 75x boost. A retweet scored 20 points. We built strategies around those specific ratios.

The January 2026 code marks all weight values as "excluded for security reasons."

What we can see is the structure. The code confirms these positive signals are tracked: favorites, replies, retweets, quotes, shares (with separate tracking for DM shares and copy-link shares), photo expansions, clicks, profile clicks, video quality views, dwell time, and author follows.

The negative signals remain severe: not interested, block author, mute author, and report. The code shows these subtract from scores, but the exact penalties aren't disclosed.

What the Code Confirms

Some patterns are clear from the source code, even without the specific weights.

Replies remain a distinct, high-value signal. They have their own dedicated tracking (reply_score) separate from likes. The infrastructure treats them as fundamentally different actions, not just "engagement" lumped together.

Author interactions still carry special weight. The follow_author_score signal suggests that when someone you don't follow creates content good enough to make you follow them, that's a strong positive signal for the content. When the original poster replies back to you, the system recognises that as meaningful conversation.

Dwell time is tracked in two different ways. The code shows both dwell_score and dwell_time as separate signals, suggesting the algorithm cares both about whether you paused and for how long. Content that holds attention gets rewarded in multiple dimensions.

The 50/50 content split persists. The architecture still draws roughly half your feed from accounts you follow (via Thunder, their in-memory post store) and half from accounts you don't follow (via Phoenix retrieval).

Author diversity is actively enforced. A dedicated scorer prevents any single account from dominating your feed, even if their content would otherwise rank highly. The system wants variety.

What This Means for Strategy

The fundamentals haven't changed. Replies still outweigh likes. Conversation still beats passive consumption. Consistent, quality engagement still compounds over time. This is why replies remain the fastest path to growth.

But the era of optimising for specific multipliers is over. You can't target a 75x boost when you don't know if that number still exists in the new system.

What works now is what's always worked underneath the numbers: creating content that generates genuine engagement. The Grok-based transformer is predicting what you'll find valuable based on patterns in what millions of users have found valuable before. When you create content that actually merits engagement, you're aligned with what the algorithm is optimising for.

The shift from hand-engineered features to pure machine learning means the algorithm is learning relevance patterns that humans might never have explicitly coded. It's watching how long people look at content, whether they expand images, whether they click profiles, whether they share privately versus publicly. Every interaction becomes a training signal.

Practical Applications

Keep prioritising replies. The code structure still treats them as high-value signals distinct from passive engagement. The hierarchy remains even if the specific ratios are hidden. Master the anatomy of a high-value reply to maximise this signal.

Focus on conversation depth, not just engagement volume. Reply chains, author interactions, and quote tweets that generate discussion are tracked as separate signals. A post that sparks ten back-and-forth exchanges likely outperforms one with fifty isolated likes.

Dwell time matters more than impressions suggest. The algorithm knows whether someone scrolled past in half a second or stopped to read for fifteen. Strong openings that stop the scroll aren't just good writing practice; they're directly rewarded by the system. See optimal posting times to maximise initial visibility.

Avoid triggering negative signals. Content that generates "not interested" clicks, mutes, or blocks isn't just ignored. Those actions actively subtract from your score. Controversial content that alienates part of your audience may hurt more than it helps, even if it generates high raw engagement numbers.

Looking Forward

𝕏 open-sourcing their algorithm twice in three years suggests this transparency will continue. The move to Grok-based transformers positions them to iterate faster, learning from engagement patterns rather than manually tuning weights.

The code also shows infrastructure for features that haven't fully materialised yet: separate tracking for different share types, product surface embeddings that record where you saw content, and video duration thresholds that hint at different treatment for short versus long video.

For now, the core strategy remains what it's been: replies over likes, conversation over broadcasting, quality over volume. The specific numbers may be hidden, but the hierarchy is confirmed in the code structure itself. For the full picture, see how the X algorithm works.

The algorithm is trying to show people content they'll engage with. When you create that content consistently, the system learns to distribute it more widely. That feedback loop hasn't changed, even if the underlying architecture has. Focus on building a brand that outlasts algorithms rather than chasing tactical optimisations.