Alibaba Releases Open-Weight Wan 2.2 Video Models With Noise-Routed MoE Architecture · cho.sh
Alibaba Releases Open-Weight Wan 2.2 Video Models With Noise-Routed MoE Architecture
AI video model releases
Alibaba Releases Open-Weight Wan 2.2 Video Models With Noise-Routed MoE Architecture
Alibaba released Wan 2.2 open-weight video models for text-to-video, image-to-video, and combined text/image-to-video generation under an Apache 2.0 license, with weights on Hugging Face and ModelScope.
The 27-billion-parameter MoE versions activate 14 billion parameters per token and route denoising between two experts: one for high-noise inputs that establishes objects and positions, and one for lower-noise inputs that adds detail.
The 5-billion-parameter Wan2.2-TI2V-5B accepts up to 512 text tokens and/or 1280x704 images, generates up to five seconds of 1280x704 video at 24 fps, and runs on consumer GPUs.
Wan2.2-T2V-A14B and Wan2.2-I2V-A14B generate up to five seconds of 1280x720 video at 30 fps; the text model accepts prompts up to 512 tokens and the image model accepts 1280x720 inputs.
Alibaba reported Wan2.2-T2V-A14B scored 85.3 for aesthetic quality on its undisclosed Wan-Bench-2.0 benchmark, above Seedance 1.0's 84.3, but scored 73.7 for video fidelity versus Seedance's 81.8.
cho.sh