Wan2.1

What it is

Wan2.1 is a video generation model from Alibaba, known in Chinese as Tongyi Wanxiang (通义万相). It makes video from a text description or from a still image. Its code and weights were published on 25 February 2025 under the Apache 2.0 license, so anyone could run the models locally.^[1]

Origins

The first models were text-to-video T2V-14B at 480P and 720P, the smaller T2V-1.3B at 480P, and image-to-video I2V-14B at 720P and 480P.^[1] Later additions were FLF2V-14B, which generates a video from a first and last frame and was added in April 2025, and VACE-1.3B and VACE-14B, all-in-one models for creation and editing added in May 2025.^[1][2]

The technical report followed on 21 March 2025.^[1]

Main features

Wan2.1 is a diffusion transformer. A diffusion model starts from noise and removes it step by step until a clean video appears. The team trained it with flow matching, a training method that teaches the model the path from noise to data. Three components work together: a 3D causal video VAE called Wan-VAE, which compresses video into a smaller representation; a T5 encoder, which reads multilingual text; and cross-attention in every block, which connects the text to the video being generated.^[1]

The sizes differ in scale. The 14B model has 40 layers, 40 attention heads and a hidden size of 5,120. The 1.3B model has 30 layers, 12 heads and a hidden size of 1,536.^[1] The 1.3B model needs only 8.19 GB of VRAM, which places it within reach of consumer graphics cards.^[1]

The models can render Chinese and English text inside the generated video.^[1] To evaluate them, the team used 1,035 internal prompts across 14 dimensions and 26 sub-dimensions, and also made manual comparisons with open and closed models.^[1]

Place in development

Wan2.1 was succeeded by Wan2.2 on 28 July 2025.^[3]