Alibaba Open-Sources Wan 2.2, First MoE Video Generation Model, With 5B Version Running on Consumer GPUs
AI video model releases
Alibaba Open-Sources Wan 2.2, First MoE Video Generation Model, With 5B Version Running on Consumer GPUs
Alibaba open sources Wan 2.2, a video generation family including text-to-video (T2V-A14B), image-to-video (I2V-A14B), and a unified 5B model (IT2V-5B).
The T2V and I2V models use a Mixture of Experts (MoE) architecture, the first in video generation diffusion models, with 27B total parameters and 14B active parameters, splitting work between high-noise and low-noise expert models for layout and detail refinement.
The MoE design cuts computational resource use by about 50% at the same parameter scale while improving complex motion, character interaction, and visual quality.
The 5B unified model uses a high-compression 3D VAE with a 4x16x16 compression ratio (64x information compression), the highest reported among open-source models, and runs on a single consumer GPU with only 22GB VRAM to generate a 5-second 720p, 24fps video within minutes.
Wan 2.2 includes a cinematic aesthetics control system letting users specify lighting, color tone, and composition keywords (e.g. sunset, rim light, low angle) to steer visual style, and the models are available on GitHub, HuggingFace, ModelScope, Alibaba Cloud BaiLian API, and the Tongyi app.
cho.sh