Kling O1

Kling O1, launched by Kuaishou on December 1, 2025, is built on a diffusion transformer that is aligned with a vision-language model, so visual and text tokens are processed in one shared embedding space. Kuaishou calls this the Multimodal Visual Language (MVL) framework, a multimodal transformer with built-in comprehension and long context that reads text, images, video clips, and specific subjects as one combined instruction.

Instead of separate models for separate jobs, O1 folds reference-based generation, text-to-video, start and end frame generation, video inpainting (adding or removing content), video modification, style re-rendering, and shot extension into a single engine. Clips run 3 to 10 seconds, and the companion O1 image model accepts up to 10 reference images for generation and detail editing. Kuaishou says the model tracks the identity of characters, props, and settings independently so they stay consistent across camera moves.

Training runs in stages: large-scale pretraining on text and video pairs, supervised fine-tuning on reference-to-video and image and video editing tasks, then reinforcement learning with Direct Preference Optimization targeting motion dynamics and visual integrity. A distillation step cuts inference from 150 to 10 function evaluations, and a cascaded multimodal super-resolution module sharpens high-frequency detail in the final output.

There are acknowledged limits. The technical report states that some controllable generation and reasoning-enhanced features are not yet available in the online version, and O1 has no native audio generation (audio lives in Kling 2.6 and 3.0, not O1). On safety, Kuaishou describes content filtering that excludes NSFW material from training data, but the report does not detail a public watermarking standard such as C2PA.