Kling 3.0 API

Kling 3.0 is the new flagship video model from Kuaishou's Kling AI, and it changes the unit of work from a clip to a scene. Instead of one continuous take, a single generation can be planned as a sequence of shots, each with its own framing, camera move and action, and the dialogue, sound effects and ambience come out of the same pass as the picture. Clips run anywhere from 3 to 15 seconds, characters can speak five languages, and the model is built to keep faces, products and on-screen text stable from the first frame to the last.

Kuaishou announced it on 5 February 2026 under the slogan of an era "where everyone can be a director", and the pitch holds up better than most launch slogans. Kling 2.6 brought native audio to the line two months earlier; 3.0 adds the storytelling layer on top: multi-shot planning, longer and more flexible durations, more languages, reusable characters and better text.

Within weeks of launch, Kling 3.0 at 1080p took first place in text-to-video on the Artificial Analysis Video Arena, both with and without audio. On Unicorn API it is a single model, kling-3.0/video, covering text-to-video and image-to-video with optional start and end frames, sound, storyboards and reference subjects.

Kuaishou, Kling and the road to 3.0

Kling is the AI video business of Kuaishou Technology, the Beijing short-video company. It launched in June 2024, and at the 3.0 announcement Kuaishou said Kling served more than 60 million creators, had produced over 600 million videos and worked with more than 30,000 enterprise clients. Kuaishou also reported that Kling's annualised revenue run rate reached 240 million US dollars in December 2025, so this is a model with a large paying business behind it rather than a research demo.

The version history is fast. Kling 1.0 arrived in mid-2024 and 1.6 at the end of that year; 2.0 and Kling 2.1 followed in spring 2025, then Kling 2.5 Turbo in the autumn. December 2025 brought two branches at once: Kling O1, a unified multimodal model that accepted text, images and video together for generation and editing, and Kling 2.6, the first Kling to generate sound with the picture. Kling describes 3.0 as an upgrade of both, built on what it calls "a deeply integrated unified model training framework".

Like every Kling model, 3.0 is closed. Kuaishou has not published weights, a parameter count or a technical report for it. It is available through Kling's own apps, the Kling API and partners such as Unicorn. At launch Kling gave early access to its Ultra subscribers first, with public availability following.

What changed from Kling 2.6

Kling's own comparison table is blunt: multi-shot, element reference, handling of three or more characters, multilingual speech, dialects and accents, 15-second output and flexible duration are all things 2.6 did not have.

Multi-shot scenes in one generation

The headline feature is multi-shot generation, which Kling markets as a built-in AI Director. In automatic mode the model reads a prompt and decides where to cut, handling patterns such as shot-reverse-shot dialogue on its own. In custom mode you write the storyboard: for each shot you set its duration, shot size, perspective, what happens and how the camera moves. Kling supports up to six shots in one clip. The practical effect is that a two-person exchange, a product reveal or a short action beat can come back already edited, with the model working to keep the space and the characters continuous across the cuts.

From 10 seconds to 15, in any whole second

Kling 2.6 offered 5 or 10 seconds. Kling 3.0 generates anything from 3 to 15 seconds in whole-second steps. Short durations make quick tests possible, and 15 seconds is enough for a complete dialogue exchange or a full social ad.

Native audio in five languages

Sound is generated with the picture: dialogue with lip sync, sound effects and ambience. Where 2.6 spoke Chinese and English, 3.0 adds Japanese, Korean and Spanish, plus American, British and Indian English accents and five Chinese dialects (Northeastern, Beijing, Taiwanese, Cantonese and Sichuanese). Kling also improved what it calls multi-character coreference: in scenes with three or more people you can specify exactly which character delivers which line.

Characters and objects that stay put

Kling 3.0 adds element references. You build an element from two to four photos of a character, product or prop, or from a short video of a character, in which case Kling extracts both the appearance and the voice tone. Elements can be combined with a start frame, so a scene can open on a specific image while a recurring character keeps their face and voice.

Text that survives the motion

Kuaishou says the model can "retain or generate text, such as signage, captions, and branded elements, with high accuracy", and that it recognises text in uploaded images, such as signs, captions and logos, and preserves it in the generated video. For product and retail work, where a garbled label ruins a shot, this is one of the more practical upgrades.

How Kling 3.0 works

Kuaishou has published little about the internals. Kling describes 3.0 as a single model trained with a unified framework that merges visual and audio generation "into a single, native stream", and that brings text-to-video, image-to-video and reference-to-video in one model. Earlier Kling models are diffusion transformers paired with an in-house 3D VAE, and 3.0 builds on 2.6 and O1, but how large the model is and what changed in the architecture Kuaishou has not said. Kling has also not published a frame rate for 3.0.

Kling 3.0 specs on Unicorn

Choose the workflow with mode (t2v or i2v), pass image and optionally end_image for animation, and set quality to std, pro or 4K. Turn on audio with sound. For a storyboard, set multi_shots to true and list the shots in multi_prompt, each with its own prompt and duration; the shot lengths must add up to the clip's duration, and storyboards take a start frame but not an end frame. Set sound explicitly on storyboards too, because its default changes when multi_shots is on. Subjects go in kling_elements, each with a unique name, a short description and its photos or video.

Where Kling 3.0 shines

Short scenes rather than single shots. Dialogue with coverage, a product demo that cuts from wide to close-up, a trailer beat: the multi-shot planner turns these into one generation instead of several clips to stitch.

Camera motion. In its launch review, Curious Refuge wrote that Kling "consistently delivered strong cinematic camera motion" and scored its image-to-video 8.1 out of 10, "easily the highest-scoring AI video model we've reviewed to date".

Multilingual speech. Five languages, accents and dialects, with control over who speaks, make localised ads and international short drama practical without a separate dubbing step.

Branded and retail content. Element references for products, plus legible signage and labels, suit e-commerce clips and ad variants.

Arena standing. In blind voting on the Artificial Analysis Video Arena, Kling 3.0 at 1080p took the top spot in text-to-video with and without audio, ahead of Grok Imagine, Runway Gen-4.5 and Veo 3.1, and placed fourth in image-to-video with audio.

Limits to know before you build on it

Fifteen seconds is the ceiling. Anything longer is assembled from several generations, and continuity between them is your job, helped by elements.

Lip sync and voice cloning are uneven. Curious Refuge found that the lip sync "doesn't always hit the mark" and that facial likeness drifted when cloning a character. Audio is strongest in English and Chinese.

Cuts are not always seamless. Testers saw colour grading shift between shots, and multi-shot work still "requires heavy iteration". An automatic multi-shot request can also come back as a single take.

Consistency across generations is not free. Kling's consistency work targets a single clip; appearance can still vary between separate generations unless you anchor characters with elements or a start frame.

Physics edge cases. Complex physical interactions may not look fully natural, so check contact and collisions closely.

Three aspect ratios. 16:9, 9:16 and 1:1 only, and with a start image the picture's own shape largely decides the frame.

Prompting tips for Kling 3.0

Write like a director. Name the shot size and the camera move, describe the action in the order it happens, and put spoken lines in quotation marks with the speaker clearly identified. Describe the action before the line, so the model ties the voice to the right person. Introduce your main characters early and describe them the same way in every shot. Clean, front-facing, well-lit reference photos on plain backgrounds give the best elements.

A single-shot prompt with dialogue:

Close-up of an old lighthouse keeper in a yellow oilskin coat, rain streaming down the window behind him, a storm lamp swinging overhead. He looks into the camera and says in a calm, gravelly voice, "She always holds. Every storm." Wind howling, waves crashing below.

A custom storyboard, written as two entries in multi_prompt:

Shot 1 (2 s): wide shot of a small tea shop on a rainy street at dusk, a wooden sign above the door reads "MOON TEA". Shot 2 (3 s): close-up of the owner pouring steaming tea into a white cup, slow push-in, warm lamp light.

A multi-character exchange with sound:

Two-shot in a cramped spaceship cockpit, red warning lights pulsing. The pilot, a young woman with a shaved head, grips the controls and says, "We have ten seconds." The engineer beside her, an older man with a grey beard, flips a switch and answers, "Make it eight." Cut to a close-up of the engineer's hand on the glowing lever. Alarms beeping, engines roaring.

Draft at Standard quality and short durations while you work out the blocking, then render the keeper longer and in Pro or 4K.

Kling 3.0 against other video models

Kling 2.6 is the direct predecessor: native audio, but only 5 or 10 seconds, no storyboard, no element references and Chinese and English speech only.

Kling 2.5 Turbo Pro is a silent option with fixed 5- or 10-second clips, good for motion tests where sound and cuts do not matter.

Wan 2.6 from Alibaba also reaches 15 seconds and is worth comparing for prompt-led shots.

Seedance 1.5 Pro from ByteDance generates silent clips or clips with audio, and is worth comparing for single shots that do not need a storyboard.

Grok Imagine Video runs longer, up to 30 seconds, when length matters more than direction.

The rule of thumb: choose Kling 3.0 when the brief is a scene, with cuts, dialogue or a character that must stay on-model, rather than a single pretty shot.

Getting started

Open Kling 3.0 in the studio, write a prompt or add a start frame, and begin at Standard quality with a short duration. For code, create a key on the API keys page and submit a job to /v1/jobs with model set to kling-3.0/video, a prompt, and the options above. The API tab on this page has full request examples, and the docs explain how jobs and polling work.

Verdict

Kling 3.0 is the release where Kling stops thinking in single clips. Multi-shot storyboards, sound in five languages with control over who speaks, 15-second clips, reusable characters and legible text make it one of the most directable video models available, and its first-place arena showing at launch suggests the picture quality kept pace with the features. Lip sync, voice cloning and continuity between cuts still need checking before anything ships. But for short scenes with dialogue, ads with cuts and characters that must stay recognisable, Kling 3.0 is the Kling model to start with.