Kling 4.0 API

Kling 4.0 is the fourth generation of Kuaishou's video model, and its pitch fits in the slogan of its launch film: "You Call the Shots". Where earlier Kling models let you describe a scene and hope, 4.0 hands over the timeline. A single generation runs up to 30 seconds, twice the old ceiling, and you can pin up to ten keyframes along it, so the opening image, the midpoint and the final pose are all yours to set. On top of that sits a reference system that takes up to fifteen inputs at once: photos, video clips, characters and voices.

Kuaishou announced the model on 28 September 2026 and opened an early Flash version to its top subscribers the same day; the full model launched on 15 October. The release targets three things Kuaishou names directly: realism, control and professional quality. In practice that means steadier motion in fast action and complex camera moves, two-channel stereo sound with better lip sync, a 21:9 ultrawide format and, in Kling's own apps, 4K output with 10-bit HDR.

On Unicorn API Kling 4.0 arrives as two models that split the release by how you steer it. kling-4.0-reference-to-video is built around the fifteen-input reference system and also runs from a prompt alone. kling-4.0-keyframe-to-video is built around first, middle and last frames. Both make clips of 3 to 30 seconds at 720p or 1080p.

Kuaishou, Kling and the road from Kling 3.0

Kling is the AI video business of Kuaishou Technology, the Beijing short-video company, and it has been one of the fastest-moving lines in AI video since its launch in June 2024. Versions 1.0, 1.6, 2.0 and 2.1 came within a year; 2.5 Turbo followed; and in December 2025 Kling split into two branches. Kling 2.6 added native audio to the classic text- and image-to-video models, while Kling O1 introduced the "Omni" idea of one engine that accepts text, images, video and subjects together.

The branches met again in February 2026. Kling 3.0 was the prompt-led model, with multi-shot clips, native audio and a 15-second limit. Kling 3.0 Omni was the reference-led one, with up to seven references, characters that bring their own voice, and editing of existing footage. Through the spring both gained native 4K, and Omni got an editing upgrade in June.

Kling 4.0 folds those ideas into one generation and raises the limits on each. The reference system is now called Omni Reference and is part of the core model rather than a separate branch. Kling has also grown as a company: two works made with Kling AI won Gold Lions at the 2026 Cannes Lions festival, the first edition to judge AI work, and Kuaishou has filed plans to spin Kling off as a separately listed company in Hong Kong.

Like every Kling model, 4.0 is closed. Kuaishou has published no weights, no parameter count and no technical report for this generation. It is available through Kling's apps, the Kling API, Kling's MCP server and command-line tool for AI agents, and partners such as Unicorn.

What changed in Kling 4.0

Thirty seconds in one take

Kling 3.0 stopped at 15 seconds. Kling 4.0 generates up to 30 seconds natively in a single pass, in any whole number of seconds from 3. That is long enough for a continuous tracking shot through a location, a complete dialogue exchange or a full product story without stitching clips together and hoping the lighting matches. Kuaishou has also announced video extension that chains several extensions up to two minutes, described as coming soon.

Ten keyframes instead of two

Previous Kling models accepted a start frame and an end frame. Kling 4.0 accepts up to ten keyframes placed along the timeline. Each one fixes a moment: a character's pose, a change of scene, a product turned to face the camera. The model fills in the motion between them. For anything with a planned structure, such as a storyboarded ad, a transformation or a choreographed move, this turns the model from a generator into something closer to an animation tool.

Omni Reference: fifteen inputs at once

Kling 3.0 Omni took up to seven references, or four when a video was involved. Kling 4.0's Omni Reference takes up to fifteen in total: up to ten images, up to five videos with a combined length of 30 seconds, and up to seven subjects, of which three can be built from video. Voices can be supplied as references and tied to a character. Kuaishou says the model reads a wide range of reference types: subject photos, three-view character sheets, grid images, storyboards, wireframes and depth-map videos.

Steadier motion and camera work

Kuaishou's headline quality claim is motion. It says 4.0 keeps large-scale movement steadier, holds continuity through high-speed action and complex camera work, and understands the standard camera vocabulary better: push-ins, pull-outs, pans, trucking, tracking and orbit shots. The company frames this as an answer to one of the most common faults in AI video, footage that wobbles or melts when the camera moves fast.

Stereo sound and more languages

Audio moves from one channel to two-channel stereo, and lip sync is described as more accurate. Speech now covers Portuguese, German and French alongside Chinese, English, Japanese, Korean and Spanish, with Cantonese and Sichuanese dialects and American, British and Indian English accents.

Ultrawide, on-screen text and HDR

21:9 joins 16:9, 9:16 and 1:1 as an output format, which matters for anything meant to look like cinema. On-screen text is supported in the core languages and can now include emoji. Kuaishou also lists 10-bit HDR output at 1080p and 4K for better highlight and shadow detail; partner documentation describes HDR as rolling out after launch, and it is not an option on Unicorn.

Longer prompts

Prompts can be far longer than before, up to 8,000 characters on Unicorn, which leaves room for a shot-by-shot script with dialogue. Kling's own examples still favour short, plain prompts; length is there for when the scene needs it.

What Kuaishou has said about how it works

Kuaishou has not published an architecture or training description for Kling 4.0. Its earlier technical report on the Omni engine described a multimodal language model that interprets mixed instructions in front of a generator working on visual and text tokens together, and Kling's earlier models are diffusion transformers. How much of that carries into 4.0, how large the model is and what frame rate it renders at are unpublished. What can be observed is the interface: references and keyframes are addressed by name inside the prompt, which tells you the model treats prompt text and visual inputs as one instruction.

The two Kling 4.0 models on Unicorn

In the reference model, every input gets a handle you use in the prompt: @image_1, @image_2 and so on for image_urls, @video_1 for video_urls, and @name for anything in elements or voices. Reference videos must be 2 to 30 seconds each and, together with video-based subjects, no more than five clips and 30 seconds in total. Without a reference video, native audio is always generated; with one, keep_original_sound decides whether the clip's own sound is kept. If you leave out aspect_ratio, the model picks automatically when references are present and uses 16:9 when they are not.

In the keyframe model, frames are addressed as @first_frame, @keyframe_1 to @keyframe_8 and @last_frame, and you may only mention frames you actually send. Duration caps how many frames you can use in total: four at 3 seconds, rising by one per second to the full ten at 9 seconds and longer. With no frames at all, the model works as plain text-to-video.

Where Kling 4.0 shines

Long continuous shots. Thirty seconds in one take changes what a single generation can be. A walk through a market, a car chase that stays coherent, or a presenter delivering a complete line all fit without a cut.

Planned structure. Keyframes make it practical to follow a storyboard. Set the opening frame, the moment the product appears and the final pose, and the model handles the motion in between. Up to ten fixed points also keep a long clip from wandering away from the plan.

Consistency from rich references. Fifteen inputs mean a character sheet, a location, a product, a motion reference and a voice can all be in the same request. For series, short drama and brand work, fewer things are left to chance.

Dialogue. Stereo sound, more languages and accents, and lip sync that early testers say largely fixes the drift seen in long dialogue on earlier versions make talking scenes far more usable.

Camera moves. Orbits, tracking shots and fast action are where Kuaishou focused its motion work, and testers report noticeably better spatial consistency than Kling 3.0, with faces holding steady across scenes.

Limits to know before you build on it

It invents details. Early testers saw the model add things that were not in any reference, such as a jacket a character never wore. When a product or costume must match exactly, check every take.

Long clips can drift. The risk of decoherence grows with length and complexity, and testers found performances could turn emotionally flat late in long sequences. Thirty seconds is possible; it is not always the best choice.

No 4K or HDR on Unicorn yet. The two models here render 720p or 1080p. Kling's 4K and 10-bit HDR output is not exposed as an option.

Strict prompt references. Every @ mention must match an input you actually supplied, or the request is rejected. That is precise, but it means templates need care.

Nothing published about internals. No parameter count, frame rate or benchmark figures came with the launch, so quality claims rest on Kuaishou's description and on testing.

Prompting tips for Kling 4.0

Write in shots and beats, not adjectives. Name the camera move, describe the action in order and put dialogue in quotation marks with the speaker identified. In the reference model, give each input a role in plain words ("@image_1 is the main character, @image_2 is the kitchen"). In the keyframe model, describe what happens between frames, because the frames already show what things look like.

A text-only dialogue scene with sound:

Two old friends lean on the railing of a ferry at dusk. The man in the grey coat says, "You kept the old camera?" The woman laughs and answers, "It still takes the best pictures." Slow push-in, gulls and engine hum in the background.

A reference-led product scene:

@image_1 is the runner and @image_2 is the shoe. The runner laces the shoe on a wet city street at dawn, then sets off. Low tracking shot alongside her feet, then an orbit as she stops at the bridge. Keep the shoe's colours and logo placement exactly as in @image_2.

A keyframe sequence:

Start on @first_frame, the empty workshop. The carpenter enters and lifts the plank to match @keyframe_1, planes it as sawdust drifts through the window light, and ends on @last_frame, the finished chair. One continuous handheld shot.

Draft at 720p and short durations while you settle timing, then render the keeper at full length and 1080p.

Kling 4.0 against other video models

Kling 3.0 remains a good choice for prompt-led clips up to 15 seconds with multi-shot cuts, and Kling 3.0 Omni for editing existing footage, which 4.0 does not offer on Unicorn.

Seedance 2.5 from ByteDance is the natural side-by-side test. Early comparisons favour Kling 4.0 on single-pass length and reference count, and Seedance 2.5 on motion physics, so run both when a shot depends on how things move and collide.

Vidu Q4 Reference To Video and HappyHorse 1.1 Reference To Video are the other reference-driven options worth running on the same character sheet.

PixVerse V6 Transition handles a simple start-to-end-frame morph; when you need frames in between, Kling 4.0's keyframe model is the step up.

Getting started

Open Kling 4.0 Reference To Video in the studio or Kling 4.0 Keyframe To Video, keep the defaults of 5 seconds at 720p, and try a prompt before adding references or frames. For code, create a key on the API keys page and submit a job to /v1/jobs with model, a prompt and the inputs you need: image_urls, video_urls, elements and voices for references, or first_frame_url, keyframe_urls and last_frame_url for keyframes. The API tab on this page has full request examples, and the docs explain how jobs work.

Verdict

Kling 4.0 is the release where Kuaishou's video model stops being a shot generator and becomes a directing tool. Thirty-second takes, ten keyframes and fifteen references are not small increments; together they let one generation follow a plan. Stereo sound, wider language coverage and steadier camera work make the result more usable straight away. The caveats are real: details the model invents, drift in long complex clips, and 4K and HDR that are not available here yet. But for anyone who has been stitching 10- and 15-second clips together, Kling 4.0 is the most controllable Kling yet.