Kling 4.0 opens to all users with 30-second AI video and MCP access

What Kling announced on 15 October

Kling AI, the video-generation service built by the Beijing-based short-video company Kuaishou, posted on its official X account at 15:00 UTC on 15 October 2026 that Kling 4.0 is now fully available. The post says the model can be used from a refreshed creation interface on kling.ai and, for the first time, through MCP and a CLI, two ways of calling it from outside the website.

That ends a staged rollout. Kling first described the model in a blog post dated 30 September, which said Kling 4.0 was entering early access with a wider rollout coming in October, and that the lighter Kling 4.0 Flash would reach Ultra yearly subscribers from 28 September. A long Kling 4.0 user guide, dated 14 October, now sets out the launch specifications in detail.

The Chinese tech site IT Home reported the full launch on 16 October and listed the same two caveats Kling did: 10-bit HDR output at 4K and 1080p is coming soon, and an API for business customers and developers will open in the near future. So this is a full release of the creative tool, not yet of the developer platform.

What Kling 4.0 can do

The headline capability is length. Kling 4.0 generates up to 30 seconds of video in a single generation, adjustable in one-second steps from 3 seconds, with 10 seconds as the default. Kling pitches that as room for long takes and continuous storytelling: an action, a reaction and a camera move that unfold in one clip rather than being stitched together from several short ones.

Sound is generated with the picture. Kling says the model produces stereo audio, with dialogue lip-synced to the characters, and the guide notes that audio-visual sync and an automatic storyboard are switched on by default with no separate toggles. You cannot upload background music or sound effects; the only audio input is a voice sample used to set how a character sounds. Music and effects are generated from the visuals and steered through the prompt.

Kling also claims steadier large motion and camera work, such as fast action, tracking and orbit shots, support for a 21:9 ultrawide frame alongside 16:9, 1:1 and 9:16, and on-screen text, emojis and logos that stay legible through movement. Spoken output covers Chinese, English, Japanese, Korean, Spanish, Portuguese, German, French and Hindi, with regional accents including Cantonese, Sichuanese, and American, British and Indian English. Prompts can be up to 8,000 tokens long, and the model will expand a short one by filling in scene blocking and camera moves.

How the references and keyframes work

Kling 4.0 has two main entry points. Omni Reference is the default: you give it images, videos, saved characters or objects, and voice samples, and describe what should happen. Keyframes is the other: you fix specific frames at specific moments and let the model fill in the motion between them. The guide is explicit that the two cannot be combined in one generation.

In Omni Reference, a single job can take up to 15 assets: up to 10 images, 5 videos, 7 elements (Kling's term for a reusable character or object, of which up to 3 can be built from video), and 3 voice profiles. The limits are specific. Images can be up to 30MB each; videos must be 3 to 30 seconds each, at most 300MB, and all video inputs together share a 30-second cap; a voice profile needs 5 to 30 seconds of audio; and the footage used to build a video element must be 3 to 8 seconds long.

Uploaded assets are numbered automatically, as @Image1, @Image2, @Video1 and so on, and Kling's advice is to say what each one is for, for example that one image sets the scene and another sets a character's face and clothes. If you give only the labels, the guide warns, the model guesses their roles and the result is less predictable. In Keyframes mode you can place up to 10 keyframe images to pin poses, expressions and scene changes along the timeline.

Kling 4.0 Flash, MCP and the command line

Kling 4.0 Flash is the cheaper, faster sibling for quick iteration and short clips. According to the guide's FAQ, it supports videos up to 20 seconds at a maximum of 720p and does not currently support start-and-end frames or multiple keyframes. The full model is the one to use for longer clips, higher resolution or precise control.

The more unusual part of the 15 October launch is the access route. Kling says Kling 4.0 is now reachable through MCP, the Model Context Protocol that lets AI assistants call outside tools, and through a command-line interface. In practice that means a creator could ask a chat assistant or a coding agent to generate a clip, rather than opening the website. It is the same pattern we have seen across creative tools, including Higgsfield's Katana MCP for Claude, where generation moves into the agent a person is already working in.

The redesigned creation page brings image, video and audio inputs into a single box, adds grid and list views for comparing outputs, keeps previews and edits on one timeline, and adds a canvas where assets and results sit as connected nodes. Kling also describes an agent on that canvas that can take a request, use tools and carry out steps of a project. Kling's posts do not give per-generation pricing for Kling 4.0, so check the plans page before planning a large job.

How to get better results, according to Kling's own guide

The user guide is unusually practical, and its main advice is to be concrete. Kling says to describe what you want to see rather than use abstract words like 'premium look' or 'cinematic feel', and to avoid negative instructions, which it says can increase errors. The two negatives it does recommend are 'no subtitles' and 'no background music', because the model may add both by default.

For longer clips, the guide suggests writing the prompt as a timeline: what happens from 0 to a few seconds, what the camera does next, and how the shot ends, followed by a line for dialogue, ambient sound, music and effects. Its worked examples run to hundreds of words, with costume details, camera paths and second-by-second action. It also says longer prompts are not always better: start with one sentence that covers subject, scene, action and emotion, then add only details that change the result.

Kling also promotes what it calls creative recreation: borrowing the pacing, shot composition and camera path of an existing video while swapping in your own product or characters, aimed at social media and advertising. That is useful, and it is also where rights questions start; copying the structure of someone else's footage is a legal and ethical call for the user, not something the tool decides.

Where Kling 4.0 sits in AI video

Thirty seconds in one pass, with synced sound and many references, is the competitive line Kling is drawing. Most video generators still produce short clips that have to be extended or edited together, and consistency across those joins is where results usually fall apart. Kling's pitch is that long takes, multiple references and keyframes together keep characters and products stable across a whole scene. Open models such as Kandinsky 6 Video are chasing the same goal with downloadable weights, while Kling stays a hosted service.

The MCP and command-line access matter for a different audience. Developers have been wiring video generation into agents and pipelines, and research such as Microsoft's Code Video model shows how far that can go when the generator is a callable tool. Until Kling's API opens, MCP and the CLI are the main ways to script Kling 4.0, which suits individual creators more than production systems.

Kling has also been promoting the model at film and tech events in the run-up to launch, including showing projects made with it at the Asian Contents and Film Market in Busan and an event at LA Tech Week on 14 October, according to its own posts. The full release puts its newest model in front of all its users while rivals ship their own long-clip and audio features.

What is still missing

Two launch features are not live yet. 10-bit HDR output at 4K and 1080p, which Kling pitches for colour grading high-contrast scenes, is marked coming soon in both the 30 September post and the 15 October announcement. So is the API for business customers and developers. Video extension up to 2 minutes is also marked coming soon in the September post. Anyone planning professional work around those features should wait for them to ship.

The quality claims, smoother motion, better lip-sync, stable text, are Kling's own, illustrated with its own sample videos and creator showcases. AI Lookout has not run Kling 4.0, and there is no independent benchmark in the launch material. The guide itself notes that features and parameters are subject to the actual product interface, so the limits above may change.