Kling 4.0 Officially Announced: 4K HDR, Native 30 Seconds, 10 Keyframes — Domestic Video Models Enter a “Spec War”

On September 28, Kuaishou’s Kling published the specs for Kling 4.0, with the official release slated for October. Kling 4.0 Flash opened ahead of it, available only to Black Gold annual card members, with specs trimmed to 720p, 8-bit SDR, and 3 to 20 seconds—trading image quality and duration for generation speed. Within the same model generation, the “fast enough” tier ships first while the professional specs are held back. That sequencing alone shows that video model vendors now have to serve two kinds of users at once: short-video creators who win on volume, and commercial clients whose cost hinges on a single finished cut.

A Few Sets of Numbers in the Spec Sheet

A single generation runs up to 30 seconds, with up to 10 keyframes, up to 15 multimodal references, and a prompt ceiling raised to 8,000 tokens. Kling breaks the 15 references down in fine detail: one pass can take 10 images, 5 video clips, and 7 subjects at the same time, mixing characters, props, and camera language into a single task. On the output side it supports 1080p and 4K 10-bit HDR, plus a 21:9 ultrawide aspect ratio; the official claim is that in high-contrast scenes, “backlight doesn’t blow out, night shots stay sharp, and neon doesn’t blur.” A video-extend feature is teased that can stretch a clip to roughly 2 minutes.

Audio is the other focus this time. Kling 4.0 supports high-quality audio and two-channel stereo, with lip-sync accuracy further refined; the official copy includes the line “no longer looks like it was slapped on in post.” For people making narrative content, audio-visual sync has long been video models’ weakest link: few models generate sound on their own, and even fewer nail the lip-sync.

The Three Players Are Chasing One Another’s Specs

Set those numbers against the market and it’s clear they weren’t chosen at random. ByteDance’s Seedance 2.5, launched in July this year, has drawn strong reviews and already supports native 4K on the product side, including in CapCut, while its API currently tops out at 1080p, 10-bit. MiniMax H3’s earlier public positioning was a maximum of 2K and 15 seconds per clip. Kling 4.0 pushes native duration to 30 seconds, stacks references to 15, and lifts output to 4K HDR—essentially aligning line by line with competitors’ spec sheets and then moving one notch further.

There’s an easily overlooked distinction here: “supported” on a spec sheet and “actually obtainable” in delivery are often not the same thing. Seedance 2.5 can output 4K on the product side, but its API is still stuck at 1080p; Kling 4.0’s 4K HDR is labeled “coming soon.” The ceiling creators can actually reach depends on which tier their channel provides, not on the biggest number quoted at the launch event.

In a September 30 note, Zheshang Securities ranked Kling 4.0’s highlights as “4K HDR, native 30 seconds, 10 keyframes,” arguing that its edge in professional creation will widen. That’s an optimistic call from a brokerage’s vantage point. Judging by the Shanghai Securities News reporter’s observation, however, in the comments on Kling’s official WeChat account users cared most about price and capabilities, and the initial verdict was “fairly lukewarm.” Between good-looking specs and users buying in, there’s still another layer.

The Business Beyond the Numbers

Kling’s business moves say more about the path it has chosen than the model itself does. On the evening of July 2 this year, Kuaishou Technology announced on the Hong Kong Stock Exchange that Kling AI had raised about US$3 billion in an independent round at a post-money valuation of roughly US$18 billion; the round corresponds to a 16.67% stake in Beijing Kling, while Kuaishou still holds about 68.33% indirectly, with a five-year IPO bet attached. It was the largest single financing in the global video generation space that year.

With that money in hand, Kling faces a very practical question: how long can the closed-source, premium-price route hold up. In that same month, industry debate over “open source versus closed source” visibly picked up. The closed camp (Kling, Jimeng, Sora) sells the single most reliable frame and keeps model weights private; the open camp (Alibaba’s Wan, Tencent’s HunyuanVideo, Zhipu’s CogVideoX) sells an ecosystem—weights are released, you can fine-tune, run locally, and there’s no per-call billing. The difference between the two paths ultimately lands on content companies’ ledgers: the single most important shot in a film is worth paying extra for quality, while the batch throughput of shooting dozens of episodes with the same character depends more on data control and speed.

One background worth noting is migration risk. Sora’s web experience went offline in April, and its API is scheduled to terminate on September 24. Closed-source models put a product’s lifecycle in the vendor’s hands, and all users can do is migrate along with it. As a company that just closed a large funding round and carries an IPO bet, Kling has more room to raise prices and adjust API terms than users might imagine.

Judgment Beyond the Specs

Back to what creators care about: what does generating 30 seconds in one pass actually change? The old approach was to break a film into 5-to-15-second shots, generate them one by one, then dub and stitch them in an editing app. Pushing a single segment to 30 seconds means some narratives need fewer cuts. But 30 seconds still isn’t long enough to “tell a whole story”; longer takes and continuous narrative mainly save you round trips rather than replacing storyboard thinking.

The 10 keyframes mean something similar. They let creators lock in a few visual beats that must appear, reducing gacha-style repeated regeneration. Stacking references to 15 shifts part of the character-consistency work away from prompt-writing tricks and into the model’s input capacity.

What this upgrade addresses is controllability, and controllability is precisely the threshold AI video must cross to go from “able to produce a clip” to “able to deliver a clip.” Questions beyond the numbers—price, licensing, data ownership—Kling 4.0 doesn’t answer, and they will determine who ultimately uses this model and on what projects.