xAI Ships Phoenix, Letting Anyone Run X's Feed Algorithm Locally
xAI ships the biggest update yet to X's open-source For You algorithm, adding a runnable local inference pipeline and a downloadable pre-trained Phoenix model
Takeaways
- xAI pushed the largest update yet to xai-org/x-algorithm, touching 187 files and adding 18,000+ lines of code.
- For the first time, a pre-trained mini Phoenix model (~3 GB via Git LFS) lets anyone run X's actual ranker locally with no training required.
- A new unified
run_pipeline.pyruns the full retrieval-then-ranking pipeline end-to-end, mirroring production. - New "Grox" service adds Vision-Language Model classifiers for spam detection, content safety, and post-quality scoring.
- The system has eliminated all hand-engineered features; a Grok-1-based transformer predicts 15+ engagement probabilities per post to compute a final score.
- Production model weights and exact scoring multipliers remain proprietary; independent verification of the live system is still impossible.
On May 15, 2026, xAI shipped the largest update to the repo yet , and for the first time, anyone can run X's actual For You ranker locally. Elon Musk announced it in a single post linking to github.com/xai-org/x-algorithm, and the repo quickly climbed past 25,000 stars. The single update touched 187 files and added over 18,000 lines of new code.
From black box to runnable code
X published its full feed algorithm on GitHub in January 2026 , the second time it has open-sourced its algorithm, the first being in 2023, though that version was never updated. The 2026 release covers the current Grok-powered system. The original 2023 release was a partial, curated dump that showed parts of the architecture but omitted training data, model weights, the trust and safety pipeline, and roughly 80% of production code. This May update is different in kind, not just degree.
It eliminates the credibility gap that plagued the 2023 release, where the published code diverged from what was actually running in production. By tying the GitHub push to the live rollout, X is making a verifiable commitment , one developers can hold them to on a four-week cycle. X committed to making the algorithm, including all code used to determine what organic and advertising posts are recommended to users, open source every 4 weeks, with comprehensive developer notes to help you understand what changed.
What actually shipped in this update
This release updates the For You algorithm code, including a runnable end-to-end inference pipeline alongside new components for content understanding, ads, and candidate sourcing. The six headline additions:
- Unified inference pipeline: A new
phoenix/run_pipeline.pyreplaces the separaterun_ranker.pyandrun_retrieval.pyscripts with a single entry point that runs retrieval then ranking from exported checkpoints, mirroring how the two stages are composed in production. - Downloadable pre-trained model: A pre-trained mini Phoenix model (256-dim embeddings, 4 attention heads, 2 transformer layers) is now packaged as a ~3 GB archive distributed via Git LFS, enabling out-of-the-box inference without training your own model first.
- Grox content-understanding pipeline: A new Python service called Grox runs Vision-Language Models over every post for spam, safety, "banger" quality, and reply ranking.
- Ads blending module: