快手的Kling

Kling AI is a family of generative video models, together with the web and mobile apps that expose them, built by Kuaishou Technology, the company behind the Kuaishou short-video app. Given a text prompt, one or more images, reference videos or audio, Kling produces a short video clip. Later versions also edit existing video and generate synchronized sound. Kuaishou sells access through consumer membership subscriptions and through API services for businesses and developers, and it reports Kling's revenue separately in its quarterly results, which makes it one of the few AI video products with public financial numbers.

Under the hood, Kling follows the same broad recipe as other modern video generators. Kuaishou's June 2024 launch release says the model uses a diffusion transformer (DiT) architecture. A diffusion model learns to turn random noise into a realistic sample step by step, and a transformer is the attention-based network design used in large language models; a DiT uses a transformer as the denoising network. The release also describes a self-developed 3D variational autoencoder (VAE), a network that compresses video into a smaller latent representation across both space and time so the diffusion model works on compact codes rather than raw pixels, and a full-attention mechanism that models spatial and temporal information together. Kuaishou has not published a full technical paper on the production models, so details beyond these descriptions are not public.

The product history is dense. Kuaishou announced Kling on June 6, 2024, claiming clips of up to two minutes at 30 frames per second and up to 1080p. Access was initially limited. When MIT Technology Review's Zeyi Yang tested it that month, he needed a Chinese phone number and a waitlist, his clips were capped at five seconds and appeared to be 720p, and each took about three minutes to generate. By Kuaishou's own count the model went through more than 20 iterations in its first year. Version 2.0 rolled out globally in April 2025, and the 2.1 series arrived at the end of May 2025 with a 720p standard mode, a 1080p high-quality mode and a premium Master edition.

The next steps added sound and unified inputs. In early December 2025 Kuaishou released Kling Video 2.6, which generates audio and video together, and Kling Video O1, which it describes as a unified multimodal model for generation, editing and video referencing, with a subject library to keep characters and objects consistent. On February 5, 2026 it launched the Kling 3.0 series, covering video and image models with Omni variants. Kuaishou's Q1 2026 report describes 3.0 as an all-in-one framework that accepts and produces text, images, audio and video, supports clips of up to 15 seconds and adds storyboard control. In Q2 2026 it added Kling 3.0 Turbo and native 4K output. fal, a model hosting company, reported that a Kling 4.0 Flash preview opened to a limited group on September 28, 2026, with the full 4.0 model set for October 2026 and generations of up to 30 seconds.

The business numbers explain Kuaishou's attention. The company said Kling passed USD 100 million in annualized revenue run rate, defined as one month's revenue times 12, in March 2025, ten months after launch. It reported USD 240 million in December 2025, with over 60 million creators, more than 600 million videos generated and over 30,000 enterprise users. Quarterly Kling revenue reached over RMB 650 million in Q1 2026, up more than 300% year on year, and over RMB 850 million in Q2 2026, up more than 200%. Kuaishou calls Kling its second growth curve.

Compared with alternatives such as OpenAI's Sora and Google's Veo, Kling competes on the same axes: motion quality, prompt following, clip length, resolution, consistency of characters across shots, and native audio. Direct, independent comparisons are hard because vendors release hand-picked demos and models change every few months. The Technology Review test in 2024 noted exactly this problem, contrasting Sora's curated showcase clips with Kling's first-attempt results. Kuaishou has emphasized long maximum durations and reference-driven consistency, and Technology Review suggested its large video platform could give it a training-data advantage.

Limitations are those of the whole field plus some specific to Kling. Early tests found blurry, grainy output that sometimes ignored large parts of the prompt, and an expert quoted by Technology Review noted that longer videos are more prone to inconsistency. Kuaishou has said only that it trains on publicly available data, so the training set is not documented. In that 2024 test English-only prompts did not work; version 2.0 later rolled out globally in April 2025.

In practice, using Kling requires an account on the Kling AI website or app and, for heavier use, a paid membership; businesses use the API. New capabilities can reach a limited group before everyone else, as with the early access preview of Kling 4.0 Flash. Teams can collaborate in a Team Plan of up to 15 members, according to the Q1 2026 report. Because model versions change quickly, the safest approach is to check the current release notes for duration, resolution and audio limits before planning a production around them.