Wan 2.2 A14B is a high-performance video generation model developed by Alibaba's Tongyi Lab. Released in July 2025 as the successor to the Wan 2.1 series, it is distinguished as the first open-source video foundation model to employ a Mixture-of-Experts (MoE) architecture. This design allows the model to expand its knowledge capacity and aesthetic range without increasing the computational overhead of each generation step.
The model features a total of 27 billion parameters, of which 14 billion are active during inference. It utilizes a specialized dual-expert system tailored for the diffusion process: a high-noise expert manages the initial stages of generation to establish global layout and scene composition, while a low-noise expert focuses on refining intricate details, textures, and cinematic lighting in the final stages. This architectural shift significantly improves semantic alignment and temporal consistency compared to traditional dense models.
Wan 2.2 was trained on a substantially larger dataset than its predecessor, incorporating approximately 65% more images and 83% more videos. This expansion enables superior handling of complex motions, such as fluid facial expressions and dynamic human interactions. The model also introduces a cinematic-inspired prompt system that grants users precise control over aesthetic dimensions, including camera angles, focal length, and lighting conditions.
The model series is released under the Apache 2.0 license, making it available for both commercial and research applications. It supports multiple input modalities, including text-to-video (T2V) and image-to-video (I2V), and is capable of generating high-definition video content at 720p or 1080p resolutions with native support for cinematic-grade aesthetics.