Wan2.2

What it is

Wan2.2 is a family of open video generation models, known in Chinese as 通义万相 Wan2.2. On 28 July 2025 the project released its inference code and model weights under the Apache 2.0 licence, a permissive open-source licence that allows the code and weights to be reused and modified^[1]. It is the successor to Wan2.1, which had been released five months earlier^[1].

The models in the first release

The initial release contained three models^[1]. T2V-A14B turns text prompts into video, and I2V-A14B turns a still image plus a prompt into video; both were offered in 480P and 720P versions^[1]. The third, TI2V-5B, is a unified model that accepts text and images together as input and outputs 720P video^[1].

Two experts inside the diffusion model

A video diffusion model generates footage by starting from random noise and removing it step by step until a coherent clip appears. In most such models the same network handles every step, from the first rough layout to the last fine textures. The A14B models split that work between two experts: a high-noise expert sets the overall layout during the early denoising steps, and a low-noise expert refines the details later^[1]. The two experts together hold 27B parameters, of which about 14B are active at each step^[1]. The point of this arrangement is to raise the total capacity of the model while keeping the computation per step closer to that of a much smaller network.

Training data and the smaller model

Wan2.2 was trained on 65.6% more images and 83.2% more videos than Wan2.1^[1]. The smaller TI2V-5B model takes a different approach to efficiency: it uses a high-compression variational autoencoder (VAE), generates 720P video at 24 frames per second, and can run on a consumer graphics card such as an RTX 4090 with at least 24GB of VRAM^[1]. The size difference between the two tiers shows up in hardware requirements. The repository states that single-GPU inference for T2V-A14B needs at least 80GB of VRAM, while TI2V-5B needs at least 24GB^[1].

Capabilities added after launch

The family grew after the July release. S2V-14B, a speech-to-video model, was added on 26 August 2025^[1]. According to the repository, support for CosyVoice text-to-speech was added to the speech-to-video task on 5 September 2025^[1]. Animate-14B, aimed at character animation and character replacement, followed on 19 September 2025^[1]. The repository also records that Wan2.2-Animate-14B was integrated into Diffusers, a widely used library for diffusion models, on 13 November 2025^[1]. A further set of weights, Wan2.2-Animate-2-14B, was published on Hugging Face in July 2026^[2].

Where it sits

Wan2.2 follows Wan2.1 by roughly half a year and keeps the same open-weights approach, publishing code and weights together rather than offering the models only through a hosted service^[1]. Two design choices define it. The first is the mixture-of-experts split in the A14B models, which treats the early and late stages of denoising as different jobs and assigns each to its own expert^[1]. The second is the range of sizes: a 5B model that fits on a single consumer GPU alongside 14B-class models that need data-centre hardware for single-GPU inference^[1]. Between them the releases cover text-to-video, image-to-video, unified text-and-image-to-video, speech-to-video and character animation, and the later ports into Diffusers put the models in front of users of the standard open-source tooling^[1].