Kuaishou rolls out Kling 4.0

Background

Generative video models turn text prompts — and often images or other reference material — into short clips. The practical limits of these systems have been fairly consistent: clips are short, movement is cautious, and holding a character or a look steady from one shot to the next is difficult. For a filmmaker, those limits decide whether a model is a storyboarding aid or something that can sit inside a real production pipeline.

What happened

On 15 October 2026 the video generation model Kling 4.0, from Kuaishou, was rolled out to all users ^[1]. The release centred on three things: longer shots, more dynamic motion, and a feature called Omni Reference ^[1].

Longer shots extend how much continuous action a single generation can hold, which matters because a sequence that has to be stitched together from many short clips is hard to keep visually consistent. More dynamic motion refers to how much movement — of camera and of subject — the model can produce. Omni Reference is named as a new capability in the release; reference-based features in generative video generally aim at letting a creator pin down an element of the output, such as a character's appearance, and carry it through generation rather than re-describing it in every prompt.

Alongside the main model, a variant called Kling 4.0 Flash was introduced, positioned around speed and volume ^[2]. Where the standard model is aimed at the best possible shot, a Flash variant is aimed at producing many shots quickly.

Response from filmmakers

Two established directors commented on the release. Lu Chuan and Li Shaohong both praised the model's instruction following and cinematic quality ^[1]. Li Shaohong singled out gains in cinematic quality and in character performance ^[1].

Instruction following means how closely the output obeys what the prompt actually asked for — the camera move, the action, the mood, the framing — rather than producing something loosely related but unusable. Character performance is the harder half: whether a generated figure reads as acting rather than merely moving. For directors, these are the qualities that determine whether a generated shot can be cut next to a photographed one.

In production

Kling 4.0 is also being used on an actual feature. Harry Hyun is using the model on the period film Natasha, which combines live action with AI-generated visual effects ^[1]. That hybrid setup — photographed material plus generated elements in the same film — is where video models are most directly tested, because generated shots have to match the look and continuity of everything around them.

Why it matters

The rollout is notable less for a single number than for where it lands. A model released to all users, endorsed by named directors for how well it follows direction, and used on a feature that mixes live action with generated effects describes a shift in how these tools are positioned: from demonstration clips towards production work.

The Flash variant points in a different direction — cost and iteration speed. In practice a production needs both: a few shots at the highest quality the model can manage, and a much larger number of cheap, fast attempts while a scene is being worked out.

The limits are the familiar ones. A general release is not the same as an independent evaluation, and a director's endorsement describes that director's experience rather than a measured benchmark. What the release does establish is the feature set — longer shots, more dynamic motion, Omni Reference, and a speed-focused variant — and the fact that filmmakers are already building it into their pipelines.