Kuaishou’s multimodal video model brings longer generation, richer references, and more keyframe control into one creative workflow.
Released October 15, 2026
Match the control to the scene: action, duration, references, keyframes, editing, or dialogue.
01 / Motion & camera
Kling describes improved handling of fast action, continuous movement, tracking shots, and camera orbits. Start from text or an image, then describe the action and the camera together. Review body structure, spatial relationships, and continuity throughout the clip.
02 / Narrative duration
The full model supports native generation from 3 to 30 seconds. Use the extra duration to connect an opening, a developing action, and a closing reaction. A longer clip gives the story room; it still needs a clear sequence.
.webp?x-oss-process=image/resize,w_1080)
03 / Multimodal references
Omni Reference accepts combinations of images, videos, subjects, and voice references, up to 15 items overall. Individual ceilings include 10 images, 5 videos, and 7 subjects. These limits share the combined allowance; they are not quantities to add together.
.png?x-oss-process=image/resize,w_1080)
04 / Visual milestones
Supply up to 10 keyframe images to establish important moments within the sequence. This adds intermediate visual guidance beyond a start-and-end pair, helping you plan subject changes, camera positions, and story progression.
Source scene
Character + setting replaced05 / Revise existing footage
Use text and visual references to revise existing footage, including subjects, backgrounds, appearance, or camera direction. The documented editing task accepts up to five clips, including one primary video, with a combined length of no more than 30 seconds.
06 / Voice & performance
Kling documents two-channel stereo output, native dialogue, and expanded multilingual support. Include who speaks, the delivery, and relevant ambience in the brief. Review the generated words and synchronization with the visible performance.
Compare the full model with the earlier generation and the lighter Flash variant.
| Dimension | Kling 4.0 | Kling 3.0 series | Kling 4.0 Flash |
|---|---|---|---|
| Starting inputs | Text, images, and supported multimodal references | Text and image generation; reference controls vary by variant | Text, image, and Omni Reference |
| Native clip duration | 3–30 seconds | 3–15 seconds | Up to 20 seconds |
| Intermediate keyframes | Up to 10 | Start and end frames; no multi-keyframe input | No first/last-frame or multi-keyframe control |
| Reference allowance | Up to 15 combined items | Up to 7 images/subjects without video; 1 video + 4 images/subjects | Omni Reference with task-specific limits |
| Resolution | 720p / 1080p / 4K listed; check mode and rollout | 720p / 1080p / 4K listed across the series | Up to 720p; 8-bit SDR |
| Workflow emphasis | Longer scenes, intermediate keyframes, multimodal edits | Shorter sequences and earlier reference workflows | Lightweight generation with a smaller control set |
Build the brief around the shot you need.
Plan the pack shot, reveal, and product use as connected visual moments. Use references for appearance and keyframes for the major beats.
Describe the performer’s movement alongside the camera direction. Review body mechanics and continuity over the full sequence.
Use the longer duration to connect action, a reaction, and a final composition, with dialogue and ambience where they serve the scene.
Kling 4.0 is Kuaishou’s multimodal video-generation model, launched on October 15, 2026. It combines longer native generation with references, keyframes, and video editing.
The full model lists 3–30 seconds per generation. Flash lists a maximum of 20 seconds. Planned extension to two minutes is separate from native clip length.
No. Flash offers a subset: up to 20 seconds and 720p, without first/last-frame or multi-keyframe controls. Confirm the model variant before planning the task.
No. The full model has a shared maximum of 15 reference items. The individual image, video, and subject ceilings do not override that combined limit.
Yes. The documented editing workflow accepts one primary video and additional supported clips, up to five total and 30 seconds combined. Inspect the revised sequence for unintended changes.
Do not assume that. Official pages contain differing rollout language. Check the current model, mode, and endpoint for the output format you need.
Kling 4.0 is Kuaishou’s multimodal video-generation model, launched on October 15, 2026. It combines longer native generation with references, keyframes, and video editing.
The full model lists 3–30 seconds per generation. Flash lists a maximum of 20 seconds. Planned extension to two minutes is separate from native clip length.
No. Flash offers a subset: up to 20 seconds and 720p, without first/last-frame or multi-keyframe controls. Confirm the model variant before planning the task.
No. The full model has a shared maximum of 15 reference items. The individual image, video, and subject ceilings do not override that combined limit.
Yes. The documented editing workflow accepts one primary video and additional supported clips, up to five total and 30 seconds combined. Inspect the revised sequence for unintended changes.
Do not assume that. Official pages contain differing rollout language. Check the current model, mode, and endpoint for the output format you need.
