Alibaba open-sources Wan2.2, a mixture-of-experts video model

Background

Wan (通义万相) is Alibaba's line of video generation models. Wan2.2 is the successor to Wan2.1, and Alibaba published it openly, so that developers and researchers can download the code and the weights and run them on their own hardware, instead of calling only a hosted service. ^[1][2]

What happened

On 28 July 2025, Alibaba released the inference code and model weights for Wan2.2 under the Apache 2.0 license. Apache 2.0 is a permissive open-source license: it allows use, modification and redistribution, including commercial use, provided the license terms and notices are kept. ^[1]

The first release covered three models. T2V-A14B generates video from text, and I2V-A14B generates video from a still image; both A14B models were offered at 480P and 720P resolution. TI2V-5B is a unified model that accepts text and an image together and produces 720P video. ^[1]

Key details

The most significant change was the architecture of the A14B models. They applied a mixture-of-experts (MoE) design to video diffusion. In a diffusion model, a clip starts as random noise that is removed step by step until it matches the prompt. In a mixture-of-experts model, several specialist sub-networks exist, and a routing step decides which of them handles a given computation, so only part of the model runs at any one time. ^[1]

In Wan2.2 the two experts take turns over the denoising steps. A high-noise expert works in the early steps, when the overall layout of the scene is being set, and a low-noise expert takes over in the later steps to refine fine detail. The two experts total 27B parameters, of which about 14B are active at each step. Parameters are the numerical weights a model learns during training, so the figure shows that the design does not run the whole model every time. ^[1]

Wan2.2 was trained on 65.6% more images and 83.2% more videos than Wan2.1. More training data of this kind is one reason a video model can keep objects and motion coherent across frames, although the release materials do not reduce the gain to a single cause. ^[1]

The TI2V-5B model takes a different route to efficiency. It uses a high-compression VAE, a variational autoencoder that squeezes video into a much smaller latent representation before the generation step, and it produces 720P video at 24 frames per second. Because of this compression, it is described as able to run on a consumer graphics card such as the NVIDIA GeForce RTX 4090, which needs at least 24GB of video memory (VRAM). ^[1]

Impact

Video generation is expensive because a clip is many frames that must agree with one another over time, so memory and compute are the main limits. Mixture-of-experts layers and compact latent spaces are two ways to reduce the work done at each step while keeping quality. Wan2.2 applied both ideas to an open model that ordinary users can run, which made the release notable for researchers who study video generation and for developers building tools on top of it. ^[1]

The release also showed that a large company can publish a video model under a permissive license, with code and weights together. For readers, the practical meaning is that such a model can be inspected, run locally and adapted, rather than being available only as a paid service.