Wan 2.7 is a multimodal AI model family developed by Alibaba's Tongyi Lab, released in April 2026. Built on a unified architecture, it integrates text-to-image generation and image editing into a single cohesive framework. The model represents a significant evolution in the Wan (formerly Wanxiang) series, focusing on resolving historical challenges such as inconsistent color reproduction and generic aesthetics through advanced semantic mapping and shared latent space processing.
The model utilizes a Mixture-of-Experts (MoE) diffusion transformer architecture, featuring a total of 27 billion parameters, with 14 billion active during any given inference pass. A standout technical feature is the integrated Thinking Mode, a chain-of-thought reasoning layer that allows the model to logically plan a scene's composition, lighting, and spatial relationships before beginning the pixel-generation process. This deliberate approach is designed to produce more coherent layouts and better adherence to complex prompts.
Wan 2.7 is available in standard and Pro variants, with the latter supporting high-definition 4K resolution output (up to 4096×4096). Key capabilities include Thousand-Face Realism for detailed facial customization and precise color control using weighted HEX values or reference palettes. The model also introduces sequential image generation, enabling creators to produce up to 12 consistent panels for storyboards or comics in a single batch. For editing tasks, it supports an interactive interface for pixel-level modifications, such as adding, moving, or aligning elements across up to nine reference images.
Official guidance for using Wan 2.7 suggests using specific technical technical descriptors rather than generic terms like "photorealistic" to achieve maximum fidelity. The Thinking Mode is particularly recommended for prompts requiring structural coherence or specific object placement. Users are also advised that the model handles camera directions in plain-language descriptions, moving away from fixed shot-type parameters used in earlier versions.