Alibaba open-sources the Wan2.1 video generation models

Background

Wan (通义万相, Tongyi Wanxiang) is the video generation line from Alibaba. Video generation models create moving footage from a prompt or from a still image, a task that needs large neural networks and a lot of computing power. Releasing the code and weights publicly lets other people run and study the models on their own hardware.

What happened

On 25 February 2025 Alibaba published the inference code and model weights for Wan2.1 on GitHub under the Apache 2.0 license.^[1]

The first release contained four models. They were text-to-video T2V-14B, available at 480P and 720P, and T2V-1.3B at 480P. It also included image-to-video I2V-14B, at 720P and 480P.^[1]

Key details

The smallest model, T2V-1.3B, needs only 8.19 GB of VRAM. This is within reach of consumer graphics cards, which is a practical point for people who want to run the model themselves.^[1]

The models use a diffusion transformer trained with flow matching. They rely on a 3D causal video VAE called Wan-VAE, and on a T5 encoder that handles multilingual text. Cross-attention in every block connects the text to the video.^[1]

The models can render Chinese and English text inside generated videos.^[1]

The team published a technical report on 21 March 2025, about three weeks after the code and weights.^[1]

Several technical terms in the release need explaining. A diffusion model generates video by starting from noise and removing it step by step until a clean sequence of frames remains. Flow matching is a training method that teaches the model a direct path from noise to real data. The Wan2.1 models were trained this way.^[1]

A VAE (variational autoencoder) compresses video into a smaller representation before the diffusion model works on it, which reduces the computing needed. The 3D causal VAE handles video so that each frame depends only on the frames before it.^[1]

The T5 encoder is a text model that turns the prompt into numbers the video model can use. Cross-attention lets the video part of the network refer to these text representations in every block.^[1]

Apache 2.0 is a permissive open-source license. Inference code is the program needed to run a trained model, which is different from the code used to train it.^[1]

Impact

The release gave developers an openly licensed video model that they could download and test. It was the first step in a series: the Wan line continued with later versions, and Wan2.2 followed in July 2025.