Exclusive Interview with Kuaishou’s Koling: Video Models Must Overcome Computing Power and Data Limitations; Google Is a Formidable Rival (AI Translation)

专访快手可灵：视频模型需突破算力和数据局限 谷歌是劲敌

The 2025 World Artificial Intelligence Conference (WAIC) in Shanghai served as a backdrop for an in-depth interview with Wan Pengfei, head of the Visual Generation and Interaction Center at Kuaishou, discussing the company's progress and strategy in AI-powered video generation. Wan noted that globally, Google remains Kuaishou’s strongest competitor due to its deep resources in AI infrastructure, talent, and the presence of YouTube. Yet, he believes the gap between Kuaishou and Google is not large, thanks to rapid advances in their own technology [para. 1].

Kuaishou’s main AI focus is not on chatbots, but rather on video generation models. While OpenAI’s Sora stunned the tech community in early 2024, Kuaishou’s KoLing—based on a Diffusion Transformer (DiT) architecture similar to Sora—was launched in June 2023. KoLing quickly moved through commercialization and built up a creator ecosystem, demonstrating Kuaishou’s deep understanding of the video market and AI’s potential to power new creative tools [para. 2][para. 3].

A key strength of KoLing is its applicability to universal video creation demands. Video is an efficient medium with vast market opportunities, and KoLing aims to enhance productivity and expand the market for both consumers and creators. KoLing’s flexibility means it’s used for both consumption and self-created content, with wide potential as long as the underlying technology continues to innovate [para. 4][para. 5].

KoLing has undergone over 30 iterations, most recently with the 2.1 series. Wan explained that KoLing 1.0 validated video generation’s practicality, while 2.0 helped standardize the industry. Rapid version cycles—every three months on average—are now a necessity in the competitive large model era [para. 6].

The model’s ability to generate local personalities and elements has kept it popular among creators, many of whom cite this as a primary reason for choosing KoLing. As of July 27, more than 45 million creators globally had used KoLing to generate over 200 million videos and 400 million images. Product development is often user-driven, with functionalities co-explored by users and engineers [para. 7][para. 8].

Scalability and data requirements remain primary challenges. Video model training needs vast and carefully structured datasets, which present higher barriers than text data. Objectively, the industry—KoLing included—has room for progress in this regard [para. 9].

Kuaishou faces competition from Bytedance, whose Seedance 1.0 Pro offers industry-low prices, and from global startups like PixVerse (over 60 million users) and MiniMax (300 million videos generated in six months), all vying for market share [para. 10].

Google’s Veo 3 is noted for its ability to generate synchronized audio with video, outpacing KoLing, which adds audio post-hoc. Veo 3’s capability to produce videos with lifelike speech and ambient sound is considered a milestone for the field [para. 11][para. 12].

AI video’s commercial value is already recognized, with adoption in film, advertising, and personal content creation, although production still requires multiple tools and manual optimization. The fundamental workflow—prompt-to-image, then image-to-video—remains unchanged [para. 13][para. 14][para. 15].

KoLing is focused on versatility, aiming to faithfully deliver user instructions without imposing its own aesthetic. Model improvements often require higher computation, influencing product pricing and optimization cycles [para. 16][para. 17].

Around 70% of KoLing’s customers are overseas, illustrating its global reach and the universality of video creation needs. Foreign users, in particular, show stronger payment willingness due to income opportunities [para. 18].

Kuaishou has prioritized its video AI initiatives, establishing an AI division in April 2024. Despite the supporting team accounting for less than 1% of Kuaishou’s 24,700 employees, KoLing achieved monthly paid revenue of over RMB 100 million in April and May. The company plans sustained, intensive investment, scaling up compute capacity to support ever-larger foundation models [para. 19].