Kuaishou announced Kling 4.0 on the last weekend of September, and the short answer to what's new is this: one generation now runs up to 30 seconds instead of 15, and you can pin up to 10 keyframes along that timeline instead of just a first and last frame. Add 15 reference slots, stereo audio with lip-sync in the same pass, 4K output and a 21:9 aspect ratio. The full model is due in October. What you can actually use today is Kling 4.0 Flash, which tops out at 20 seconds and 720p.
Length is the headline everyone repeats. The keyframes are the part that changes how you work.
Why this one matters more than a spec bump
Kling 3.0 already planned its own cuts. We wrote about that shift in the shot list moving inside the model: you describe coverage, the model decides where to cut. The weakness was control. You could set a start frame and an end frame, and everything in between was the model's opinion.
Ten keyframes turn that opinion into something closer to a storyboard. Kling's official Kling 4.0 page pitches them as anchors you place anywhere on the timeline to steer camera movement, shot changes, character entrances and reveals. If that holds up in the full release, a 30-second generation stops being a long gamble and becomes a sequence with fixed points you chose.
That matters for anyone cutting narrative work. Most scenes are not one shot. They are four or five beats that have to land in a specific order, and until now the order was the thing you fought hardest for.
What Kling 4.0 changes compared with Kling 3.0
Here is the side-by-side, based on Kling's own announcement and the spec roundup on Morphic's Kling 4 page:
- Length: 3 to 30 seconds per generation, up from a 15-second ceiling.
- Control: up to 10 keyframes, up from first and last frame only.
- References: up to 15 at once, mixing images, video clips, saved elements and voice samples.
- Prompt: up to 8,000 tokens, enough for a written shot list rather than a paragraph.
- Picture: up to 4K, a 21:9 option, and 10-bit HDR, which Kling still lists as coming soon.
- Sound: stereo audio with lip-sync generated alongside the picture.
Two things are missing from that list. Kling hasn't published per-second pricing for the full model, and nobody outside early access has run it. Treat every claim above as a claim until October.
What Flash actually delivers today
Kling 4.0 Flash opened on September 28 to Ultra Yearly subscribers first. It generates 3 to 20 seconds at up to 720p. That resolution is fine for previs and animatics, and too soft for anything you plan to project.
The early hands-on reports are useful because they are mixed. MindStudio's first look found lip-sync accuracy and spatial consistency clearly better than 3.0, with eyelines and continuity holding across the clip. It also found performances that felt emotionally flat even when every technical element worked. That tracks with what we see across models. Getting a face to stay the same is a solved-enough problem now. Getting it to act is not.
Bloomberg framed the launch as Kuaishou chasing ByteDance, and the timing supports it. Seedance 2.5 shipped 30-second continuous output in July. Kling matching that length was expected. Pairing it with 10 keyframes is the part ByteDance doesn't have yet.
A keyframe plan you can reuse
Don't write a 30-second prompt and hope. Write the keyframes first, then the prose that connects them. Here is the plan we would use for a short scene: a courier delivers a package to a woman who wasn't expecting one.
- 0s: wide, rainy street at dusk, courier on a bike entering frame left.
- 6s: medium, courier at a red door, package under one arm.
- 11s: close, door opens, woman's face, wary.
- 17s: over-the-shoulder on the courier, he says one line.
- 23s: insert, her hands turning the package, no return address.
- 30s: close, she looks up the street. He's already gone.
Attach a character sheet for each person and a location still for the doorway as references. Keep the dialogue to one line, because two-speaker exchanges still break more often than they work. Then run the same plan on Flash at 20 seconds by dropping the 6s and 23s beats. If the four remaining anchors hold at 720p, the full model is worth your credits when it lands.
What to do about it
If you make narrative work, start thinking in keyframes now, even on older models. The habit costs nothing and transfers directly. If you make ads or product loops, wait for 4K and published pricing before you rebuild a pipeline around it. If you only have Flash access, use it for animatics and timing tests, not deliverables.
At Promvie, we already break a script into scenes and shots before anything renders, so a model that takes a timeline of fixed beats fits the way our pipeline works rather than fighting it.
Thirty seconds was always going to arrive. Being able to say where those seconds go is the actual upgrade.