Meituan's unified recommendation foundation model MTFM lifts orders up to 6.68%, cut inference cost 24%
美团技术团队 · wechat · 2026-09-22
Meituan details MTFM, a unified recommendation foundation model accepted to KDD 2026, unifying ranking across its food delivery businesses. Heterogeneous tokenizers, dynamic masking, and a hybrid Target/Full attention architecture enable multi-scene training with User-Level samples. Online results: order volume up 2.06%–6.68% across businesses, inference cost down 24%.
More from Infra
- DeltaTensors Stores Fine-Tunes as Weight Deltas: 953MB → 294MB With Minimal Quality Loss — cupheadgamer · 2026-09-22
- Alibaba unveils new AI chip it calls China's most powerful — AIFlow_ML · 2026-09-22
- What coding models and quants do you run locally on DGX Spark? — be566 · 2026-09-22
- Intel Ships Day-0 OpenVINO Support for Qwen-Image-2.1 — Alibaba_Qwen · 2026-09-22
- Acer CEO says memory prices to fall in 2027 as CXMT starts mass production — AIFlow_ML · 2026-09-22
- Alibaba roadmap: Qwen explores RSI, new models up to 10T params, 20GW datacenters by 2032 — teortaxesTex · 2026-09-22