Kling O1

What it is

Kling O1 is a video model that Kuaishou announced on 1 December 2025. The company introduced it as the "world's first unified multimodal video model" and called it the "industry's first unified multimodal creation tool"^[1]. The claim is about packaging rather than about a single feature: text, video, image and subject inputs all go into one engine, and that same engine is used both to generate footage and to edit footage that already exists^[1][2]. Clips run 3 to 10 seconds^[1].

Video tools have conventionally been split in two. Generators make a clip from a prompt or a reference image. Editors modify a clip that already exists, usually by drawing masks to mark which pixels should change, or by setting keyframes to mark when the change should happen. Putting generation and editing into one model means the same input can serve both as something to make and as something to change, and that is the property Kuaishou is advertising here.

The framework underneath

Kuaishou says the model is built on a Multimodal Visual Language (MVL) framework, which pairs a multimodal transformer with a multimodal long context^[1]. A transformer is the neural network architecture behind most current language and image models; it lets every part of the input attend to every other part, so relationships are learned from data instead of being hand-coded. Long context refers to how much input the model can hold in view at once; a longer context lets the model condition on more frames, more reference material or a longer instruction without losing the beginning. Calling the transformer and the context multimodal means they are not restricted to one data type: text, images and video frames are handled in a shared representation. That is what allows a single engine to accept a written prompt, a source clip and a photographed subject together.

What it does

The launch material lists the tasks the model covers: reference-based generation, text-to-video, start/end-frame generation, in-painting, video modification, style re-rendering and shot extension^[1].

Several of these are editing operations that would otherwise need separate tools. In-painting fills in a region of the frame. Style re-rendering restyles existing footage. Shot extension carries a clip on past its original ending. First- and last-frame generation, where the opening and closing images are supplied and the model produces what happens between them, was said to soon support the same 3 to 10 second range^[1].

Editing in words

Editing is driven by text rather than by manual masks or keyframes^[1]. Example prompts given at launch include "remove passersby", "transition day to dusk" and "swap the protagonist's attire"^[1]. None of them asks the user to trace around an object or to set a time range by hand.

The model also supports what Kuaishou calls skill combos, which run compound edits in a single pass, for instance inserting a new subject into a shot while at the same time modifying the background^[1]. A dedicated subject library is offered so that characters and objects stay consistent from one shot to the next, a feature aimed at film and television work, where continuity across shots matters^[1].

Place in the release

Kling O1 was the centrepiece of what Kuaishou called its Omni Launch Week in early December 2025. The same run included the image model Kling Image O1, Video 2.6 and Digital Human 2.0^[3]. Image O1 generates images from text or from up to 10 reference images^[1].

The "first" framing is the company's own; what the release documents is the range of inputs and the list of tasks described above.