Kling 4.0 Launches: 30-Second Video, Stereo Audio and a Flash Tier

Kling AI released Kling 4.0 on October 19, 2026. According to The Decoder's report, which cites the company's announcement, the new video model generates up to 30 seconds in one pass, accepts up to ten keyframe images, understands prompts up to 8,000 tokens, outputs stereo audio, and ships alongside a lighter Flash variant. The company's own release notes page tracks the version history.

This is the confirmation of what we flagged in September as a Discord leak, not a launch. Below: what is live, what is still "coming soon," what the first comparisons say, and what to do if you build on video models.

TL;DR: Kling 4.0 at a glance

What is actually new in Kling 4.0

Longer clips and keyframe control

Doubling single-pass length from the 15-second cap of Kling 3.0 to 30 seconds matters less for raw duration than for what it replaces: stitching. Most AI video workflows chain several short generations and then fix continuity by hand. With ten keyframes you can define character states and scene transitions up front, so one generation can carry a short scene with a beginning, middle and end.

Filmstrip of matching shapes illustrating Kling 4.0 keyframe-controlled video consistency

Prompt comprehension also grows to 8,000 tokens, enough for a shot list written as a screenplay rather than a one-line wish.

Omni Reference and Creative Recreation

The upgraded Omni Reference accepts up to 15 reference assets: up to 10 images, 5 videos with a combined maximum of 30 seconds, and 7 elements. According to The Decoder, storyboards, wireframes, image grids, white-model videos and depth-map videos all work as references. That points toward a pipeline in which a 3D blockout or a storyboard drives the final look.

A "Creative Recreation" feature takes an existing video as a template and swaps products or characters, and the company says facial expressions, body movement, camera angle or background can be changed without disturbing the rest of the clip.

Stereo audio and multilingual lip sync

Kling 4.0 adds two-channel stereo audio and lip sync across more than ten languages, including German, English, Japanese, Korean and Spanish, with dialects and accents. In-video text generation works in multiple languages and is said to stay stable through camera movement. For marketers localizing one ad into many markets, this is the most commercially useful line in the release, and also the one to test hardest, since lip sync quality varies by language.

Piano keys becoming a ribbon, representing stereo audio generation in Kling 4.0

Kling 4.0 Flash

Flash trades resolution and length for speed and cost: it tops out at 720p and 20 seconds. Third-party trackers reported a limited early-access window for Flash from late September; the full model followed in October. Use Flash for drafts and iteration and reserve the full model for finals.

MCP and CLI access

Kling 4.0 can be called through MCP and a CLI, so coding agents can script generation. If you already run agent workflows, see our guides on Claude Code video editing and agentic video production with OpenMontage for how those pipelines are structured.

What is not available yet

Kling lists three items as coming soon or unannounced:

1. API support, with pricing not announced.
2. 10-bit HDR output at 4K and 1080p.
3. Video extension to as long as two minutes.

Do not plan a production pipeline around these until they appear in the changelog. The same caution we gave in September applies: a feature on a launch slide is not a feature you can ship.

How does it compare with Wan 3.0 and Seedance 2.5?

The Decoder published a side-by-side of Kling 4.0 against Wan 3.0 and Seedance 2.5 and concluded that Kling "holds its own but doesn't pull ahead," especially on logical shot sequencing, physics and natural-sounding dialogue. The outlet floated that this might signal the pace of video-model progress leveling off, while noting that the next Wan and Seedance releases will say more.

That is a qualitative comparison from one outlet using prompts it chose, not an independent benchmark. For context on the rivals, see our posts on Seedance 2.5's 30-second 4K claims, Alibaba's HappyHorse 1.1, and the open-weight Kandinsky 6 Video.

We deliberately do not fill in rival cells that the cited sources do not give; check the linked posts for their specs.

From leak to launch: the timeline

The leaked slider range turned out to match the launch claim of 30 seconds, and the Flash name was right. What the leak did not show, and the launch adds, are Omni Reference with 15 assets, stereo audio, the MCP and CLI route and the 8,000-token prompt window.

What people are asking about Kling 4.0

Is 30 seconds really one continuous shot? The company says up to 30 seconds can be generated in a single pass. It does not say the clip is one unbroken camera take; the keyframe feature suggests multi-beat scenes with transitions. Test whether cuts, character identity and lighting stay consistent across the full length before you promise a client a half-minute scene.

Does Flash replace the main model? No. It caps at 720p and 20 seconds, so it is a drafting tool. The pattern most video teams settle on is to explore ideas cheaply, lock a prompt and keyframes, then re-render the winner on the full model.

Can I use the API today? Kling describes API support as coming soon and has not announced prices. MCP and CLI access exist, which is useful for agents and scripts, but check the terms of use and credit costs before automating at scale.

Will 4K and HDR arrive? The company lists 10-bit HDR at 4K and 1080p as coming soon. Until it lands, assume delivery at the resolutions Kling documents for the tier you use.

Is the multilingual lip sync production-ready? The company claims more than ten languages. Expect the best results in widely used languages and uneven results in dialects. Always have a native speaker review dubbed output before publishing.

A starter prompt structure for keyframed scenes

Because the model accepts ten keyframes and an 8,000-token prompt, a structured brief works better than a single sentence. A template you can adapt:

Scene: 25-second product reveal, one location, golden-hour light.
Keyframe 1 (0s): empty workbench, camera low and still.
Keyframe 5 (12s): hands place the product, camera pushes in.
Keyframe 10 (25s): product centered, soft rack focus to the logo area.
Audio: stereo room tone left, gentle foley right; one line of dialogue in Spanish.
Constraints: no text overlays, keep the product color constant across all frames.

This is our suggested format, not an official Kling recipe. The point is to describe state at each keyframe and constraints that must hold throughout, then compare the output against your brief shot by shot.

What to do now

1. Test with your own shot. Generate the same 10-second prompt on Kling 4.0, Kling 3.0 and one rival. Judge physics, hands and dialogue.
2. Try keyframes. Give it first, middle and last frames of a scene and see whether it respects the in-between.
3. Check lip sync in your languages, not just English.
4. Use Flash for iteration and keep an eye on credits, since pricing for the full model and API is not yet public.
5. Wait on the API before wiring Kling 4.0 into a product; confirm model IDs in the official changelog.

Why it matters

Video generation is moving from "wow, a clip" to controllable production tools: references, keyframes, localized audio and agent access. Kling 4.0 puts a Chinese lab at parity with its closest rivals on those axes rather than clearly ahead, which is itself a signal. For the broader landscape, read our video generation guide covering Sora, Runway and Kling and the MiniMax H3 open video model post.

Related reading

• Kling 4.0 preview leak, September 2026
• Video generation AI guide: Sora, Runway, Kling
• Seedance 2.5: 30-second 4K AI video
• HappyHorse 1.1 from Alibaba
• Kandinsky 6 Video, open source
• MiniMax H3 open video model

Specifications are accurate as of October 19, 2026, per the sources cited; Kling may change features, limits and pricing.