Technology company Alibaba has launched Wan2.2, a set of open-source AI video generation models designed to make it easier for creators and developers to produce cinematic-style videos. The models, which use a Mixture-of-Experts (MoE) architecture, can create videos from text or images with a single click.
The Wan2.2 lineup includes three models: Wan2.2-T2V-A14B for text-to-video, Wan2.2-I2V-A14B for image-to-video, and Wan2.2-TI2V-5B, a hybrid model that can handle both tasks. These models were trained on carefully selected visual data to produce videos with professional-grade aesthetics, such as control over lighting, color tone, camera angles, and composition.
The MoE models are also designed to handle complex movements, from realistic facial expressions and hand gestures to sports sequences, while keeping the visuals consistent with physical laws. To reduce the high computing power typically needed for video generation, Alibaba developed a two-step denoising process: one expert focuses on the overall scene, while another refines details. Although each model has 27 billion parameters, only 14 billion are used at a time, cutting processing needs by up to half.
Wan2.2 also features a cinematic-inspired prompt system that lets users specify aesthetic elements like lighting or composition in detail. Compared to its earlier version, Wan2.2 was trained on significantly more data—about 66% more images and 83% more videos—allowing it to produce more complex and creative results.
The hybrid model, Wan2.2-TI2V-5B, uses a compressed 3D architecture to process videos efficiently. It can create a five-second 720p video in minutes using a standard consumer graphics card, making it more accessible for independent creators.
The Wan2.2 models are available for free download on Hugging Face, GitHub, and Alibaba Cloud’s ModelScope platform. Alibaba’s earlier Wan2.1 versions have attracted more than 5.4 million downloads to date.