AI Affairs, home

Friday 2 October 2026

Technology

Kuaishou previews Kling 4.0 with 30-second clips and more controls

Kling 4.0 Flash opened to Ultra Yearly subscribers on 28 September. The full model is scheduled for October. A clip-extension feature is listed as coming soon.

A modern commercial building exterior with trees and a sidewalk in front
Photo: N509FZ, CC BY-SA 4.0, via Wikimedia Commons (cropped)

Kuaishou’s Kling AI announced Kling 4.0 on 28 September 2026, offering video generation of up to 30 seconds in one pass and more ways to direct what appears in a clip, Pandaily reported on 29 September. A lighter version, Kling 4.0 Flash, opened to Ultra Yearly subscribers on the announcement day. The full model is scheduled for October.

Key points

  • Kling 4.0’s stated single-pass limit is 30 seconds, up from 15 seconds for Kling 3.0.
  • A task can use up to 15 references, while up to 10 keyframes can guide changes within a clip.
  • Flash opened to Ultra Yearly subscribers on 28 September; the full model is due in October.
  • An extension feature intended to take clips to about two minutes is listed as coming soon.

Kling 4.0 doubles the single-pass limit

The 30-second limit doubles Kling 3.0’s 15-second ceiling. “Single pass” matters here: it describes how much video the model can produce in one generation, rather than the length of a finished piece assembled from separate clips. Kling presents the longer span as support for continuous action and a narrative within one clip, where a change in setting or camera position has to remain coherent with what came before it.

That limit is separate from a video-extension feature. Kling lists repeated extensions, which are intended to continue a clip to about two minutes, as coming soon. A 30-second generation is therefore the length offered for a single pass, while the longer duration belongs to a planned way of continuing an existing result.

The output options extend beyond duration. Kling 4.0 supports 4K and 1080p output and a 21:9 ultra-wide format, while 10-bit HDR is listed as coming soon. It also supports dual-channel stereo audio. Kling says the model handles rapid movement, tracking shots and cameras moving around a subject more steadily, and that speech is more closely aligned with lip movements.

Speech support covers Chinese, English, Japanese, Korean, Spanish, Portuguese, German, French and Hindi, along with accents and dialects including Cantonese. The model can render text in multiple languages on screen, as well as emoji and logos. These are output capabilities Kling describes, rather than results from a reported test of how reliably each appears in generated video.

Fifteen references and 10 keyframes guide a clip

In one Kling 4.0 task, the 15-reference allowance can include no more than seven subjects, five video clips and 10 images. The separate limits describe the material that can be supplied, while 15 is the overall ceiling for references in a task. Images and footage can give the model something concrete to follow for a character, a movement, a camera angle or sound.

Users can also set up to 10 keyframes to direct character states, changes of scene and plot beats. A keyframe is much like a marked picture in a storyboard: it pins down a point the clip should reach, while the generation fills the motion between marked points. That gives a more specific instruction than a written request for a scene to change at some point during the video.

Making a short product advert could begin with pictures and footage supplied as references, then use marked frames to guide when the scene changes. The longer single-pass limit would allow more of that sequence to be requested in one generation; the planned extension feature would be a separate way to continue it.

Written instructions can run to 8,000 tokens, and Kling 4.0 can expand a short prompt. Its editing controls can add, replace or remove subjects and backgrounds in existing video, or change details including weather, colour, material and viewpoint. For someone building a video workflow around the model, those controls offer distinct ways to supply source material, direct the sequence and revise a result.

A remake mode uses a reference video’s camera choices and pacing to construct a new commercial clip. Kling aims that mode at uses including social media, product advertising, fashion, beauty and consumer electronics. It is a different use of reference footage from asking the model to carry a particular subject or movement into a newly generated scene.

Kling 4.0 Flash opens to annual subscribers

Kling 4.0 Flash opened to Ultra Yearly subscribers on 28 September. Kling positions this lighter version for faster generation at lower cost and for creating video in volume. The company has also redesigned its creation page so image, video and audio references enter through one input box, with a timeline for previewing and editing and a canvas workspace that includes an agent assistant.

API pricing for the new model had not been published, Pandaily reported. That price matters to developers and companies intending to make repeated generation part of a product or production process: the subscriber opening gives access to Flash, while API pricing would determine the stated charge for building on the model through that route.

Kling’s commercial position extends beyond this release. Its annualised recurring revenue had reached approximately $500 million by March 2026, Crypto Briefing reported. The publication also reported that revenue in the second quarter of 2026 exceeded RMB 850 million, a 200% increase from a year earlier. Those figures concern the video business, rather than a measurement of Kling 4.0’s output.

Kuaishou submitted plans in May 2026 for Kling to become a separate company and seek a Hong Kong listing, Crypto Briefing reported. A financing round in July valued Kling at around $18 billion. The model announcement arrives while Kling is being developed both as a video product and as a business intended to stand apart from its parent.

Kling is scheduled to release the full Kling 4.0 model in October, Pandaily reported.

Topics: Foundation models, Inference