Wan 2.7 is a multimodal video generation model developed by Alibaba’s Tongyi Lab, released in early 2026 as a major evolution of the Wan series. It utilizes a unified architecture to handle text-to-video, image-to-video, and video-to-video editing tasks. The model is notable for its integrated native audio generation, which produces synchronized sound effects, dialogue, and background music directly within the video synthesis pipeline, ensuring tight alignment between visuals and sound.
The model is built on a 27 billion parameter Mixture-of-Experts (MoE) backbone, with approximately 14 billion parameters active during each inference step. A defining feature of Wan 2.7 is its "Thinking Mode," a pre-generation reasoning phase where the model interprets the semantic depth of a prompt and plans the scene's logical composition before rendering frames. This approach aims to minimize common generative artifacts, improving temporal coherence and physical plausibility in motion.
Wan 2.7 introduces advanced control mechanisms for creators, including first-and-last frame generation, which allows users to specify both the starting and ending states of a scene for precise narrative control. It also features Subject Referencing, allowing for the maintenance of consistent character identities across multiple clips by processing visual references. The model supports high-definition output at 720p or 1080p resolutions at 30 frames per second, with durations typically extending up to 15 seconds.
For optimal results, official documentation suggests a prompt formula consisting of Subject + Scene + Motion + Sound Description. Users can also leverage the 9-grid image-to-video workflow, which uses a structured layout of reference images to provide the model with richer spatial and compositional guidance during the generation process.