Over 18k new lines added, mini Phoenix is only 3GB, this lightweight design is quite good, want to dismantle it to see how the recall layer interacts with user behavior

X Official Algorithm Repository's second major submission was in May 2026, with 187 files, 18,263 lines added, and 926 lines deleted.
The core addition is phoenix/run_pipeline.py, which implements end-to-end inference: starting from checkpoints, user behavior, and precomputed corpus to perform recall, interaction probability prediction, and ranking score synthesis, making it closer to real-world recommendation.
Additionally, about 3GB of mini Phoenix models are used for example inference.
Ad insertions are constrained by security, risk, keyword, and brand safety considerations, with new grox content understanding pipelines covering spam detection, classification, and multimodal embedding.
Overall, it fills gaps in the recommendation peripheral production chain, though it is not complete code, but it facilitates research and disassembly.