Kling 2.6

A video model that makes its own soundtrack

Kling 2.6 is a video generation model from Kuaishou, released on 3 December 2025. Its defining feature is what the company calls "simultaneous audio-visual generation": the model produces the picture, the voiceover, the sound effects and the ambient background sound in a single pass, rather than delivering a silent clip that has to be paired with audio afterwards^[1]. Each clip can run up to 10 seconds^[1].

Generating sound and image in the same step is the interesting part. When the two are decided together, a line of dialogue and the mouth that speaks it, or a footstep and the floor it lands on, come out of one process. Split the job in two — picture first, audio later — and the sound stage has to infer what is happening on screen, then line its output up with it; when that inference is wrong, sound and picture drift apart. Doing both at once also changes the workflow: a user writes a prompt and receives a finished clip with sound, instead of a silent file that still needs to be scored^[1].

Two ways in

The model upgrades two existing modes: text-to-audio-visual, in which a written prompt yields a clip with sound, and image-plus-prompt-to-audio-visual, in which a still image plus instructions becomes a moving shot with audio^[1]. The two modes correspond to two common needs — describing a scene from nothing, or setting an existing picture in motion and giving it a voice.

On the audio side, the model generates Chinese and English voices, and the types of audio it covers include speech, dialogue, narration, singing, rap and ambient effects^[1]. So it is not only dubbing: it aims at the layers of sound a scene might contain, including several characters speaking inside one clip. Kuaishou describes the model as holding a "world-leading position in Chinese voice generation performance"^[1] — a claim a company makes about its own product, and one scoped to a particular language.

What it is aimed at

Kuaishou positions Kling 2.6 around three kinds of work: short advertisements with narration and product showcases, multi-character skits of the sort made for social media, and narration for e-commerce product videos^[1]. What these share is a need for someone to be talking while sound is present: an ad needs one voice to carry the selling point, a skit needs several characters to take turns speaking, a product video needs a voice that can sit directly on top of the footage. Multi-character dialogue is the harder part of that set, because it requires different voices within a single clip, matched to the people on screen.

The 10-second ceiling shapes what the model is for^[1]. It suits shot-level fragments rather than complete narratives; in practice a model like this is used as a material generator, producing piece by piece and leaving assembly to the edit.

Where it sits in the Kling line

Kling 2.6 belongs to the 2.x stage of the Kling video models. In the same quarter, a motion control feature was added that lets users replicate specific movements, and it drew wide attention on global social platforms^[2]. In its fourth-quarter 2025 financial results, Kuaishou reported Kling revenue of RMB 340 million^[2]. Kling 2.6 was succeeded by the Kling 3.0 series in February 2026^[2], which places it as a step in a fast-moving sequence rather than an endpoint: a release that folded sound into the generation step, and was itself replaced a little over two months later.