Kling opens 4.0 Flash early access to Ultra Yearly subscribers

• Kling opened 4.0 Flash to Ultra Yearly subscribers on September 28; the full 4.0 model is due in October, per Kling's announcement.
• Early Flash tests top out at 720p and 20 seconds, according to WordTrafficker's test and a creator's settings screenshot. The advertised 30-second, 4K feature set belongs to the broader 4.0 rollout.
• A 15-second ocean race was directed in five timed beats using one text prompt, as AllaAisling's example demonstrates.
• Character-reference sheets are already turning into dialogue scenes: ozansihay's hangar test pairs two people with a 20-second Turkish exchange.

A full heist prompt specifies which character speaks in each shot and forbids the model from treating a multi-angle character sheet as a crowd. An official feature description reproduced by IT之家 mentions a canvas Agent and a shared editing timeline alongside the headline video specs.

Kling's official announcement says Flash is available to Ultra Yearly subscribers now, with the flagship 4.0 planned for October. An early teaser said only “You call the shots” in Kling's teaser; MayorKingAI's pre-release reply held off on judging the promised longer clips before the reveal.

The early-access UI shows a 720p, 20-second, 16:9 generation with one output selected; choices for two to four outputs carry VIP badges. WordTrafficker also reported a 20-second, 720p Flash limit and said 1080p was still to come. Kling's detailed description likewise marks 1080p and 4K 10-bit HDR output as coming soon, so the flagship specs should not be read as today's Flash export options.

AllaAisling made a 15-second race clip from a text prompt that breaks the action into a shot-by-shot clock:

1. 0–3s: Two hypercars cross an exposed ocean floor at sunrise.
2. 3–6s: Water begins to reclaim the track behind them.
3. 6–9s: They accelerate toward a wall of seawater.
4. 9–12s: The cars burst through the water curtain.
5. 12–15s: They emerge onto dry track as the sea closes behind.

The prompt also calls for photorealistic water physics and golden reflections, and explicitly excludes fantasy creatures. The generated clip is attached to AllaAisling's post.

For a 20-second aircraft-hangar scene, ozansihay fed the model multi-angle sheets for two characters. The posted prompt tells it to preserve faces, clothes and tattoos, avoid turning sheet panels into extra people, cut to the correct speaker during each Turkish line, and finish by tracking backward as the woman leaves the workbench.

A separate test reused a character already saved in Elements in ozansihay's text-to-video example. Kling's 3.0 Omni guide documents Element references and multi-shot text-to-video in the preceding model, so the saved-character workflow predates 4.0.

A creator's ferry-scene prompt uses two character references, six reference images and six timed shots over 30 seconds, down to distinct Glaswegian and Texan voices. The same example thread spans several different production briefs:

• Product ad: One bottle reference image, four shots in 15 seconds, with instructions to keep the label facing camera in the bottle-ad prompt.
• Selfie vlog: A continuous 30-second, vertical take asks for digital noise, drifting framing, changing auto-exposure and compressed phone-mic audio in the walking-vlog prompt.
• Spy chase: Six cuts in 15 seconds specify eyelines, frame-right running direction and a crate spill in the night-market prompt.
• Car drifting: A multi-shot example uses only “Generate a multi-shot car drifting action scene” in the short-prompt test.

Those 30-second examples appear in a thread about 4.0's coming capabilities; they do not change the 20-second Flash limit shown in early-access tests.

Early creators are also testing styles away from photorealistic action. Artedeingenio animated one of their illustration styles in a 20-second Flash test, while koldo2k's animation post focused on multi-shot spatial and visual coherence. A liminal dreamcore test combines a woman's monologue with cuts to scenes from her past.

Artedeingenio said their Seedance prompt structure did not seem to transfer cleanly to Kling in a follow-up, and later said they needed more testing before drawing a verdict.

Kling lists these targets for the flagship model in its announcement, with additional detail in the reproduced official description:

• Story length: Up to 30 seconds in one generation and up to 10 supplied keyframes.
• References: Up to 15 multimodal inputs in a generation, including a stated ceiling of 10 images, five videos and seven subjects.
• Picture and sound: Planned 1080p and 4K 10-bit HDR output, stereo audio, improved lip sync and 21:9 framing.
• Script detail: Prompts up to 8,000 tokens, plus multilingual dialogue and accent support.

The official description also says repeat extensions could bring a video to two minutes, but explicitly labels that extension feature as coming soon.

Beyond the model, Kling's reproduced official product description says its creation page has been redesigned: image, video and audio references can share one input box; results can switch between grid and list views; previewing, cutting and adding assets happen on one timeline. It also describes a canvas where assets and outputs can be arranged as connected nodes with an Agent alongside the work. The company gives no separate rollout date for each part of that editing surface.