Kuaishou's Kling AI has announced that its next-generation video generation model, Kling 4.0, will launch officially in October. A lighter version, Kling 4.0 Flash, already opened for limited access on September 28, with slots going to Black Gold annual subscribers. Kling has not yet announced the exact launch date for the full version, its rollout scope, or pricing.
The official keywords for this generation are "realism, control, and professionalism," which break down into four areas: audiovisual performance, all-around reference support, storytelling ability, and content types.
Spec sheet
| Item | Kling 4.0 |
|---|---|
| Output format | 4K and 1080p, both supporting 10-bit HDR |
| Native generation length | Up to 30 seconds |
| Continuation extension | Up to 2 minutes |
| Reference input per job | Up to 10 images, 5 video clips, 7 subjects |
| Multiple keyframes | Up to 10 |
| Prompt length | 8,000 tokens |
The two rows with the most direct effect on how the model is used are duration and reference input.
Thirty seconds means a long shot with a full narrative arc — setup, development, twist, resolution — can be generated in one pass instead of being stitched together from several clips. Kling also says large-scale motion and complex camera moves are more stable in this generation, and specifically calls out high-speed motion shots as a target use case this time. Continuation can extend the same sequence to 2 minutes, which suits short dramas and ads that need continuous storytelling.
On the reference-input side, images, video clips and subjects can now be mixed within a single generation task. Anyone who has worked with AI video knows that a character's face shifting in the third shot is the most common reason for a redo. Kling says consistency is "more stable" in this release, but it hasn't provided benchmark data that would allow a side-by-side comparison — the real-world result will have to wait until third parties get hands-on access.
Audio and editing also get an upgrade
Beyond visuals, Kling 4.0 adds dual-channel stereo sound and improves how closely a character's mouth movements match the dialogue. On the editing side, it lists two new features: "precision editing," for making local changes to an already-generated clip, and "creative remix," which reuses the pacing and camerawork of an existing video as a template for a new one.
The 10-bit HDR support is aimed squarely at professional post-production. Standard 8-bit footage tends to band when pushed through color grading curves; 10-bit leaves far more headroom. This capability won't matter much to casual short-video viewers, but it's genuinely useful for teams pulling AI-generated footage into an editing timeline.
What domestic users can access first
On the capital side, Kling has been active lately in China's video-generation race. It closed a $2.8 billion funding round in July at a valuation of roughly $18 billion; in early September, the China AI Industry Investment Fund subscribed for about 1.14% of the company for RMB 1.4 billion, with redemption rights attached to the terms. Kling 4.0 is the first major version release since both of those rounds closed.
For creators in China, Kling remains one of the few first-tier video models that can be used directly without workarounds, unlike overseas options such as Runway or Veo. For now, only Black Gold annual members can access 4.0, and what they get is the Flash version; official materials don't spell out the differences in duration or resolution between Flash and the full release.
Among comparable products, MiniMax's Design, released in August, takes a different approach: 15-second single shots stitched into a full sequence by a separate tool. Kling instead pushed the length of a single generation itself to 30 seconds. Which approach suits longer content better depends on whether characters and scenes can stay consistent across a long shot — which is exactly the metric Kling hasn't published numbers for.
Sources: Kuaishou Kling AI official announcement, Jiemian News, CocoLoop, IT Home, 21st Century Business Herald; output specs, input limits and access scope verified against Kling's official announcement.