OneRec
End-to-end generative recommender that replaces the retrieve-then-rank cascade; +1.6% watch time on Kuaishou.
About
OneRec is Kuaishou's answer to the cascaded recommender stack: instead of separate retrieval, ranking and re-ranking stages, a single encoder-decoder model generates a whole session of recommended short videos from the user's behavior history. Items are tokenized into hierarchical semantic IDs, capacity is scaled with a sparse mixture-of-experts, and the model is aligned with iterative preference optimization (DPO adapted for recommendation).
The technical report claims it handles 25% of Kuaishou's query traffic with App Stay Time up 0.54% (and 1.24% on Kuaishou Lite), at operating costs of just 10.6% of the traditional pipeline. A follow-up, OneRec-V2, moved to a fully autoregressive decoder with beam search. Like HSTU, the production model isn't downloadable — the value here is the published architecture and the proof that end-to-end generative recommendation works at hundreds-of-millions-user scale.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (arXiv 2502.18965): +1.6% watch time in Kuaishou main-feed deployment; ~1B MoE model, session-wise generation
- Technical report (arXiv 2506.13695): 25% of QPS, App Stay Time +0.54%/+1.24%, OPEX 10.6% of traditional pipeline, 23.7%/28.8% MFU
- OneRec-V2 (arXiv 2508.20900): +0.467%/+0.741% App Stay Time in A/B tests
- No independent replication — Kuaishou-internal metrics
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: two arXiv technical reports (2502.18965, 2506.13695) plus OneRec-V2 (2508.20900) with production A/B metrics; widely cited in 2025 recsys surveys. Strengths: rare public documentation of a production generative recommender at massive scale. Docked for: (1) all metrics are Kuaishou-internal and not independently verifiable; (2) no public code or weights. License: not applicable — no public release.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 3.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · ArXiv reports 2502.18965 / 2506.13695 / 2508.20900 (V2); +1.6% watch time Kuaishou main feed; 25% of QPS; App Stay Time +0.54%/+1.24%; OPEX 10.6% of traditional pipeline
- · Widely cited in 2025 recsys surveys; rare public documentation of production generative recommender at scale
- · All metrics Kuaishou-internal, not independently verifiable; no public code or weights