HSTU
Meta's generative recommender architecture that reframes ranking as sequential transduction — with the first demonstrated recsys scaling law.
About
HSTU — Hierarchical Sequential Transduction Units — is the architecture at the heart of Meta's 2024 paper 'Actions Speak Louder than Words', which proposed treating recommendation as a generative modeling problem instead of a scoring problem. Rather than extracting features for every (user, item) pair, HSTU merges a user's entire history into one chronological sequence and predicts the next action autoregressively, unifying retrieval and ranking in a single model.
Meta reported that a 1.5-trillion-parameter HSTU-based recommender improved metrics 12.4% in online A/B tests across surfaces with billions of users, and — the headline for researchers — demonstrated a scaling law for recommendations across three orders of magnitude, the way LLMs scale. Meta open-sourced the modeling code (meta-recsys/generative-recommenders) and NVIDIA published training/inference examples, but the production weights stay internal, so outsiders can study the architecture, not deploy Meta's model.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (arXiv 2402.17152): 1.5T-parameter HSTU recommender, +12.4% online metrics in Meta A/B tests; scaling law demonstrated across three orders of magnitude
- Secondary reporting (Synerise explainer): 5.3–15.2x efficiency vs FlashAttention-2-style Transformer++
- Adopted as an MLCommons training benchmark reference
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: peer-reviewed ICML 2024 paper (arXiv 2402.17152) with Meta-scale online A/B results; open modeling code (facebookresearch/generative-recommenders); adopted as MLCommons training benchmark; NVIDIA added HSTU attention kernels to cuDNN. Strengths: landmark paper, broad industry corroboration, code is genuinely open. Docked for: (1) production weights are not public — 'open' applies to code only; (2) the 12.4% gain is a Meta-internal metric, not independently reproducible. License: code under Meta's open-source terms (see repo).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · ICML 2024 paper arXiv 2402.17152; 1.5T-param recommender, +12.4% Meta online A/B metrics; first recsys scaling law across 3 orders of magnitude
- · Open modeling code (facebookresearch/generative-recommenders); MLCommons training benchmark; NVIDIA cuDNN HSTU attention kernels
- · Production weights not public — 'open' applies to code only; 12.4% is Meta-internal metric