Kuaishou’s AI XiaoKuai: Technical Innovations Behind the Consumer‑Entertainment Assistant

AI XiaoKuai is Kuaishou’s official consumer‑entertainment assistant built on the company’s self‑developed large‑model platform “Kuaishou Yi”. It supports video Q&A, knowledge Q&A, and exhibits strong personality traits, engaging users in fun conversations and gaining 10 million new followers within six months.

1. Birth and Development of Large‑Model Technology Language has always been the bridge for human‑machine interaction. After the 2018 introduction of the Transformer architecture, researchers created BERT with a masked language modeling objective, achieving significant gains on GLUE. In 2020, OpenAI released GPT‑3 with 175 billion parameters, followed by ChatGPT, which applied instruction fine‑tuning (SFT) and reinforcement learning from human feedback (RLHF) to dramatically improve instruction following and dialogue capabilities. The rapid release of dozens of domestic large models in 2023 marked a new competitive era for AI.

2. Technical Innovations of the Kuaishou Yi Model Kuaishou built a trillion‑parameter‑level training and inference infrastructure, optimizing MFU to industry‑leading levels. Leveraging massive short‑video and live‑stream data, the team curated high‑quality multimodal tokens for pre‑training. After a year of development, Kuaishou launched the “Kuaishou Yi” model (KwaiYii) in three sizes—13B, 66B, and 175B—each with a base version (KwaiYii‑Base) and a chat version (KwaiYii‑Chat). The 175B model approaches GPT‑4 performance on benchmarks such as MMLU, C‑Eval, GSM‑8K, and HumanEval.

3. AI XiaoKuai – A Multimodal, Personified Companion Robot The assistant combines multimodal video understanding, retrieval‑augmented generation, and community knowledge to answer factual and video‑content questions with usefulness, fun, and warmth. Its personable replies (e.g., “the sweetest thing in the world may be love or friendship”) have attracted massive user engagement, contributing to a rapid 10 million‑follower increase.

To improve long‑turn dialogue, Kuaishou introduced two key technologies: the Parrot user‑question simulator for generating extensive multi‑turn data, and DialogBench, a comprehensive long‑dialogue evaluation suite covering intent detection, slot filling, knowledge, commonsense, and persona perception across 12 tasks. Both works received high‑score paper acceptances at ACL’24 and NAACL’24.

AI XiaoKuai was also recognized in the “AIGC Best Practice Top 20” by InfoQ’s AIGC Pioneer List for its scenario innovation, practical results, and industry value.

Recruitment Notice Kuaishou is hiring for multiple positions related to large‑model research and multimodal AI, including senior algorithm engineers, AIGC application experts, and internship roles. Interested candidates can submit resumes via the provided email addresses or QR‑code links.