Kuaishou

corporate

china · Founded 2011-03 · IPO 2021 (HKEX: 1024) · Market cap: $26B (HKG: 1024) as of 2026-03

Kuaishou (HKG: 1024) is China's second-largest short video platform after Douyin, founded in 2011 by Su Hua and Cheng Yixiao. The company went public on the Hong Kong Stock Exchange in February 2021 in one of the year's largest tech IPOs. Tencent is the largest institutional shareholder (~17%), but founders retain control via a dual-class voting structure.

Kuaishou's AI ecosystem is characterized by high-performance multimodal models decentralized across several teams. The Kling Team (led by Kun Gai) produces Kling, a video generation model that gained significant traction in 2024 as one of China's leading text-to-video systems. The Kolors Team develops image generation models, and the Kwaipilot Team (led by Zizheng Zhan) focuses on coding and reasoning models including KAT-Coder and Klear.

As a platform serving over 600 million monthly active users, Kuaishou applies AI extensively to recommendation, content understanding, livestream commerce, and real-time video effects. The company's competitive "frenemy" dynamic with Tencent — receiving WeChat ecosystem access while competing directly with WeChat Video Accounts — shapes its strategic positioning.

Website Wikipedia HuggingFace GitHub Artificial Analysis
multimodalgenerationvideocodingrecommendation

People

Outputs (19)

Kling 3.0 2

model paper

Most advanced Kling model, supporting native 15-second high-fidelity clips and multi-shot storyboarding. Includes Kling-Omni unified video generation/…

KuaiSearch

dataset

Large-scale e-commerce search dataset covering recall, ranking, and relevance.

OpenOneRec

library

Open framework and benchmark for generative recommendation based on the Qwen3 backbone.

↓ 147

KAT-Coder 2

model paper

72B active parameter MoE coding model using the "AutoThink" paradigm. Designed for agentic coding with parallel tool calls to reduce interaction laten…

Klear 3

model paper

Sparse MoE LLM family from Kuaishou, including a 46B-parameter base model (2.5B active) and the Klear-Reasoner variant achieving 90.5% on AIME 2024 vi…

↓ 2.3k

KAT (Kwai-AutoThink) 2

paper model

KAT (Kwai-AutoThink) series for dynamic fast/slow reasoning. Includes the foundational research report and an open-source experimental coding variant.…

↓ 138

Kwai Keye-VL 2

model paper

8B parameter multimodal foundation model specializing in short-video understanding, with accompanying technical report.

↓ 79.9k

SRPO: Staged History-Resampling Policy Optimization

paper

Two-stage reinforcement learning framework for LLMs that surpasses DeepSeek-R1-Zero-32B on AIME24 and LiveCodeBench with only 1/10 of the training ste…

Kling 2.0 1

model

Major update to the Kling video generation family, introducing simultaneous audio-visual generation and "Elements" for character/object consistency.

KuaiFormer

library

Transformer-based retrieval framework for real-time industrial content recommendation.

Kolors 2

model paper

Large-scale text-to-image diffusion model based on the General Language Model (GLM), optimized for high-quality Chinese and English text rendering.

↓ 594

LivePortrait 2

library paper

Efficient portrait animation framework using implicit keypoints for stitching and retargeting control. Achieves 12.8ms generation speed on RTX 4090. T…

↓ 5.1k 📄 9

Kling 2

model paper

Original video generation model utilizing Diffusion Transformer (DiT) architecture, capable of 2-minute videos at 1080p/30fps.

Direct-a-Video: User-Directed Camera Movement and Object Motion

paper

Research on user-directed camera movement and object motion. Accepted at CVPR 2024.

KwaiYiiMath 2

model paper

Specialized reasoning variant fine-tuned for mathematical problem-solving, with accompanying technical report.

KwaiYii

model

General-purpose LLM series with sizes ranging up to 175B parameters.

KuaiSAR

dataset

Unified search and recommendation dataset containing genuine user behavior logs.

KuaiRand

dataset

Unbiased sequential recommendation dataset with millions of "randomly exposed" interactions.

KuaiRec

dataset

Fully-observed dataset for recommender systems with near 100% density.

📄 8