Alibaba says RecGPT-V3 cuts serving cost 52.4% while lifting Taobao GMV 3.97%
burny_tech · x · 2026-07-27
Alibaba’s RecGPT-V3 technical report says LLMs can act as the “brain” of large-scale recommender systems.
- The system is a stateful, hybrid-modal recommender with a Memory Hub for long-horizon user history, plus a Hybrid-modal Foundation Model that reasons over text tags and Semantic IDs.
- It reduces user-modeling computation by 55.8% and lowers output token cost by 200× through latent reasoning.
- Deployed on Taobao’s “Guess What You Like” feed, it reportedly improves IPV by 1.28%, CTR by 1.00%, TC by 1.97%, and GMV by 3.97%, while cutting end-to-end serving resource consumption by 52.4%.
- The paper frames RecGPT-V3 as an attempt to overcome three recommender bottlenecks: stateless reprocessing, tag-to-item information loss, and inefficient explicit reasoning.
Related event: Alibaba's RecGPT-V3 Cuts Taobao Compute Costs and Boosts GMV(2 posts)→
More from Companies & People
- OpenAI co-hosted GPT-6 hackathons in SF and NYC with Cerebral Valley — OpenAIDevs · 2026-09-23
- Meta's Alexandr Wang reveals muse has been in the works since at least Sept 2025 — adrianscottcom · 2026-09-23
- OpenAI shares GPT-6 hackathon tale: builder used real star maps with Astra to find way home — OpenAIDevs · 2026-09-23
- Instacart lands on Meta's Muse: say "Taco Tuesday" and get an auto-built grocery cart — alexandr_wang · 2026-09-23
- Databricks ships GPT-6 Sol/Luna and Claude Opus 5.5 with Unity Gateway model governance — matei_zaharia · 2026-09-23
- Fields Medalist Martin Hairer Explains Why AGMAI Is Truly Independent of OpenAI — AlexKontorovich · 2026-09-23