Unveiling NetEase’s ‘YuZhi’ Multimodal Model: Boosting Personalized Recommendations

01: Behind Personalized Recommendation – Multimodal Large Model

NetEase’s Fuxi research unit launched the “YuZhi” multimodal understanding model in 2021. Built on an image‑text dual‑tower architecture, it has been trained with 200 M, 400 M and 900 M parameters. Using NetEase’s open‑source EET inference framework, the model is compressed and optimized for hardware, achieving a four‑fold speedup and lower deployment cost.

In zero‑shot evaluations on business datasets, “YuZhi” outperforms Chinese‑CLIP (CN‑CLIPV iT‑H/14). The model now powers personalized recommendation in NetEase News—providing image‑text representations for a dropout‑net recall optimizer—and in NetEase Cloud Music, where it enhances content representation, cold‑start retrieval and CTR prediction.

02: NetEase + Ascend AI – Creating >1+1 Intelligent Experiences

Huawei’s Ascend AI built a full‑process large‑model enablement platform, released in 2022. With MindSpore’s automatic mixed‑parallel API and high‑level Transformer API, developers can implement trillion‑parameter models with minimal code while achieving superior performance.

The platform provides fine‑tuning toolkits that support few‑shot training and one‑click model adaptation, already used in the Zidu TaiChu open‑service platform. Its model‑compression utilities cut compute by about 70 % with negligible accuracy loss and boost inference speed by over 20 %.

Collaborating with NetEase, the teams combined “YuZhi” with Ascend AI’s capabilities to create the “YuZhi‑Wukong” image‑text understanding model. Deployed in NetEase’s video recommendation pipeline, it improves core algorithm metrics by roughly 5 % and is slated for broader use in News and Cloud Music.

Future plans include extending these multimodal models to more NetEase services and partnering with additional industry players to advance large‑model innovation across the AI ecosystem.