Large-Scale Deep Learning Systems and Their Application at Xiaohongshu (RED)

AI-driven information technology is ushering in a new wave of scientific advancement. Xiaohongshu, one of the fastest‑growing mobile‑internet platforms in China, has built a massive UGC community centered on image‑text and short‑video content. The platform generates enormous multimodal data and user‑behavior feedback daily, creating both valuable opportunities and technical challenges.

At the REDtech Youth Technical Salon on October 15, Xiaohongshu’s Vice President of Technology, Kaichi, presented “Large‑Scale Deep Learning System Technology and Its Application at Xiaohongshu,” unveiling the internally developed LarC platform.

Kaichi’s background includes a degree from Shanghai Jiao‑Tong University, former VP of Technology at YY, and chief architect at Baidu’s Fengchao, where he worked on CTR machine‑learning algorithms for search advertising.

Business Overview

Xiaohongshu now has over 200 million monthly active users, with more than 70 % born after 1990. The community emphasizes authentic, everyday life sharing by ordinary users, which drives high trust in content and fuels a “grass‑planting” (种草) commercial model that converts content exposure into brand impact.

Technical Challenges

The platform deals with massive multimodal data (images, text, video, behavior logs) at a scale of hundreds of billions of feedback samples per day. Key research problems include extracting user‑interest signals from this data, real‑time personalized recommendation, and handling long‑tail content diversity.

Real‑Time Recommendation System

The recommendation pipeline uses the LarC machine‑learning framework to train multi‑task models that predict clicks, dwell time, likes, and saves. Parameters are managed by a large‑scale, conflict‑free parameter server.

Online training works as follows: user interactions (browse, click, like) are streamed through Flink, assembled into high‑performance samples, and fed to the model for inference. The same samples are used for a short‑interval online training step that updates model parameters within minutes, after which the refreshed model is deployed instantly.

To address recommendation “filter bubbles” and short‑term freshness, Xiaohongshu employs diversified sequence modeling and replaces traditional DPP diversity with an SSD algorithm that efficiently slides windows over the feed, leveraging Siamese neural networks to assess long‑tail content similarity.

The team’s work on sequence‑level value estimation and diversity was published at KDD 2021.

Multimodal “Life‑Search” Engine

Given the richness of lifestyle information on the platform, Xiaohongshu built a next‑generation multimodal search engine that understands both visual and textual cues, enabling personalized searches such as “how to style a dress for different occasions.” This relies on large‑scale representation learning using models like T5, BERT, and GPT trained on the platform’s multimodal corpus.

Generative AI for Native Commercial Content

To help merchants create native‑style commercial posts that blend seamlessly with user‑generated content, the team applies large‑scale generative models to suggest titles and copy tailored to Xiaohongshu’s tone. Experiments show improvements in click‑through rate, dwell time, and overall user satisfaction.

LarC – The In‑House Large‑Scale Machine Learning Platform

LarC was initiated in 2019 and expanded to search, recommendation, and advertising by 2021, becoming a fully platformized solution in 2022. It now covers infrastructure, compute frameworks, resource scheduling, offline training, and online deployment, enabling AI engineers to process massive datasets and train large deep‑learning models efficiently.

Overall, the presentation highlighted how Xiaohongshu leverages cutting‑edge AI, multimodal learning, and large‑scale system engineering to power its rapidly growing content ecosystem.