Kling 2.5 Turbo API

Kling 2.5 Turbo is the new video model from Kling AI, the generative video business of the Chinese short-video company Kuaishou. It turns a text prompt, or a still image used as the first frame, into a 5 or 10 second clip at up to 1080p, and it arrives with an unusual combination of promises for a model called "Turbo": better quality than Kling 2.1, not just faster output.

The quality claim has outside support. Within ten days of launch, Kling 2.5 Turbo took first place in both the text-to-video and the image-to-video rankings of the Artificial Analysis Video Arena, a leaderboard built on blind votes between clips, moving ahead of Hailuo 02 Pro, Google's Veo 3 and Luma's Ray 3.

What changed is mostly about motion and obedience. Kling 2.5 Turbo follows longer, multi-step prompts more faithfully, moves its subjects and its camera with more energy, and keeps the look of a source image when it animates it. On Unicorn API it is available as Kling 2.5 Turbo Pro, with text-to-video and image-to-video in one model.

Kuaishou and the Kling line

Kling went into public testing in June 2024 inside KuaiYing, Kuaishou's video editing app, and has been updated at a steady clip since: Kling 1.6 at the end of 2024, 2.0 in April 2025, and 2.1 in May, offered in Standard, Pro and Master quality modes. Kuaishou says Kling has passed more than 30 iterations, generated over 200 million videos and 400 million images, and serves more than 20,000 enterprise customers and a community of over 45 million creators.

Kling 2.5 Turbo was announced on 23 September 2025. Like every Kling model it is closed: Kuaishou has not released weights, a parameter count or a technical report, and the model is available only through Kling's own web and mobile apps, the Kling API and partners such as Unicorn. Kuaishou's head of operations, Zeng Yushen, framed the release around where these tools are heading: "AI is now primarily a tool, but in the future, it will become a true partner."

What changed from Kling 2.1

Prompts with more than one step

Kuaishou's headline upgrade is prompt adherence. The company says the model now handles "complex multi-step instructions and causal relationships", and that users can steer how a clip progresses, how characters interact and how a scene changes through the prompt alone. In practice that means a prompt can describe a sequence, such as a runner accelerating, leaping a gap and landing in a roll while the camera follows, and the model should play the beats in order rather than blur them into one movement.

Bigger, faster motion

The second upgrade is motion. Kuaishou promises "more dynamic motion and camera movement with greater amplitude": fights and sprints with a tracking camera, and complex action such as figure skating, synchronised swimming and group dance. The company says it used reinforcement learning to simulate real-world physics, so that high-speed scenes such as skateboarding or artistic swimming play out more smoothly.

Image-to-video that keeps the style

For image-to-video, Kuaishou says 2.5 Turbo captures the artistic style of the source "even for nuanced details including colors, lightings, textures, and overall atmosphere". That matters for anyone animating illustrations, concept art or paintings, where the point of the clip is that it still looks like the original artwork once it moves. The company also cites improved lighting and composition across both modes.

What "Turbo" means

The "Turbo" part of the name points to speed, but Kuaishou has not published a figure for how much faster generation is than on Kling 2.1. The measurable changes in this release are the ones above: prompt adherence, motion range and style preservation.

How Kling 2.5 Turbo is built

Kuaishou has said little about the internals of 2.5 Turbo. The architecture it has described for Kling in general is a diffusion transformer paired with an in-house 3D variational autoencoder that compresses video across space and time, with full attention to model motion and scene changes. Beyond the mention of reinforcement learning for physics, the training changes behind 2.5 Turbo, its size and its training data are not public.

Kling 2.5 Turbo specs on Unicorn

Upstream, Kling also supports an end frame for image-to-video; Unicorn exposes the first frame only.

Benchmarks and early reception

Kuaishou's own blind tests report win-loss ratios against rival models: in text-to-video 285% against Seedance 1.0 mini, 212% against Veo 3 Fast and 160% against Seedance 1.0; in image-to-video 208%, 289% and 164% against the same three. A ratio of 160% means Kling's clip was preferred about 1.6 times as often as it lost. These are the maker's numbers, but the independent Artificial Analysis arena pointed the same way, with Kling 2.5 Turbo at the top of both of its video leaderboards in early October.

Early hands-on reviews were positive with clear caveats. Tom's Guide called it "a giant leap forward" and "a very impressive tool", praising a folding-city tracking shot and a superhero transformation that kept tattoos and lighting consistent on the costume. Curious Refuge scored it 7.4 out of 10 and singled out its visual fidelity as "some of the most consistent and convincing" it had seen, with "sharper motion, stronger lighting" and grounded light that makes clips look as if someone actually shot them. Its summary of the model's comfort zone: "one character, doing one action, in one environment".

Where Kling 2.5 Turbo shines

Action and sport. Sprints, spins, jumps, skateboard runs and dance routines are the material Kuaishou built this release around, and its showcase clips pair them with a camera that tracks the subject. When the scene has one athlete and one clear move, this is where the model's wider motion range pays off.

Cinematic single-subject shots. Reviewers single out its lighting, and Kuaishou cites better composition. A single person or product in a well-described environment, with one clear action and one camera move, is where the model looks most like real footage.

Animating artwork. Style preservation in image-to-video makes it a strong choice for bringing illustrations, paintings and concept frames to life without losing brushwork, palette or texture.

Prompts with a sequence. Because it handles causal, multi-step instructions better, you can script a small story beat in one clip, an action and its consequence, rather than generating each moment separately.

One model for both modes. Text-to-video and image-to-video share one model with the same controls, so a pipeline can switch between generating a shot from scratch and animating a supplied frame by changing a single mode parameter. That keeps batches of social clips, ad variations or previz shots simple to run.

Limits to know before you build on it

Silent clips. Kling 2.5 Turbo generates no audio. Music, dialogue and sound effects need a separate tool and a separate pass.

Crowds and acrobatics. Temporal consistency falters when several characters share the frame, according to Curious Refuge, and faces can drift. Tom's Guide found the model could not complete a somersault: it "forgets what my face looks like" mid-action and repeats a jump instead of finishing the flip. Fine detail can wobble too, with fingers that "occasionally bend or lengthen subtly", and wide landscape shots can show slightly unnatural parallax.

Ten seconds, three shapes. Each generation is 5 or 10 seconds, and text-to-video offers only 16:9, 9:16 and 1:1. Longer pieces have to be stitched from several clips.

Strict filters. Kling operates under Chinese content rules that block politically sensitive subjects, and one review of 2.5 Turbo summed it up as "way more censored" than before, with some prompts rejected before anything renders. Allow time for rephrasing if your subject matter sits near the edges.

No end frame on Unicorn. Image-to-video takes a first frame only, so you cannot pin both ends of a transition in one generation.

Prompting tips for Kling 2.5 Turbo

Structure the prompt in five parts: the subject, the action and its pacing, the camera move, the environment and lighting, then the style and mood. Use specific motion verbs ("pivots", "vaults", "glides") and name one primary action per clip. In a 5-second clip, avoid stacking several extreme camera moves, which tends to produce disorienting motion, and keep complex multi-character choreography for longer clips or separate shots. Use the negative prompt for artefacts you keep seeing, such as blur or distorted hands. Raise guidance_scale toward 1 when the model ignores part of your instructions, and lower it to give the model more creative latitude.

An action shot with a tracking camera:

A figure skater in a deep red costume glides across a frozen mountain lake at sunrise, accelerates, launches into a high spinning jump and lands cleanly, spraying fine ice crystals. The camera tracks alongside at ice level, then swings around her as she lands.

A multi-step prompt with cause and effect:

A courier on a bicycle races down a rainy city street at night, swerves around a delivery van, hits a puddle that sends a wave of water over the pavement, and a pedestrian jumps back laughing. Low tracking shot, neon reflections on wet asphalt.

An image-to-video prompt for a painting, describing only what moves:

The waves roll gently against the boats, lantern light flickers on the water and the clouds drift slowly. Slow push-in toward the harbour. Keep the thick oil-paint brushwork and warm colours of the painting.

Kling 2.5 Turbo against other video models

Kling 2.1 Master was the top tier of the previous generation. With the newer model ranked higher in blind tests, there is little reason to choose 2.1 Master for new work. Kling 2.1 Pro (1080p image-to-video) and Kling 2.1 Standard (720p image-to-video) remain options for existing pipelines, and on Unicorn 2.1 Pro accepts an end image, which 2.5 Turbo does not, so it is still the Kling to use when a shot has to land on a fixed final frame.

Hailuo 02 Pro from MiniMax is one of the models Kling 2.5 Turbo overtook in the Artificial Analysis arena, but arena rankings average over many prompts, so a side-by-side test on your own subject matter is worthwhile.

Seedance 1.0 Pro from ByteDance renders at 720p or 1080p, and its strength is multi-shot storytelling within one clip. Kuaishou's own tests put Kling 2.5 Turbo ahead in head-to-head preference; Seedance is the one to try when a clip needs several shots rather than one continuous move.

Getting started

Open Kling 2.5 Turbo in the studio, write a prompt and render a 5-second test before committing to 10 seconds. For code, create a key on the API keys page and submit a job to /v1/jobs with model set to kwaivgi/kling-v2.5-turbo-pro, plus prompt, duration and aspect_ratio. For image-to-video, set mode to i2v and pass the first frame in image. The API tab on this page has full request examples, and the docs explain how jobs and polling work.

Verdict

Kling 2.5 Turbo is a clear step up from Kling 2.1. It follows complex prompts more faithfully, delivers larger and more convincing motion, and keeps an image's style when animating it. The independent arena results back up Kuaishou's own claims, putting it at the top of both text-to-video and image-to-video rankings at launch. It is still silent, still capped at 10 seconds, still uneasy with crowds and acrobatics, and strictly filtered. For single-subject action, cinematic product and lifestyle shots and animated artwork, it is the Kling to use.