Alibaba Wan

History

2025: Product Announcement

The Chinese company Alibaba has completely opened access to its Wan 2.1 neural network, designed to generate realistic images and videos. This became known on February 26, 2025. The model is distributed with open source and is available to users free of charge for both academic and research tasks and commercial use.

According to Reuters, the Wan 2.1 model, first introduced in January 2025, occupies a leading position in the VBench rating, which assesses the quality of generative video models. Wan 2.1's ability to process scenes with the interaction of several objects is especially highly appreciated, which makes the created content as realistic as possible.

𝓲

Фото: Freepik

Alibaba has released four variants of the Wan 2.1 neural network, differing in functionality and requirements:

1. T2V-1.3B - a text-video model with 1.3 billion parameters, optimized for work on home video cards, requiring 8.19 GB of video memory to generate five-second video in 480p resolution in four minutes on the RTX 4090 GPU;
2. T2V-14B - an improved text-video model with 14 billion parameters, supporting video generation in 480p and 720p resolutions based on descriptions in Chinese and English;
3. I2V-14B-480P - image-video model with 14 billion parameters for creating videos in 480p resolution based on uploaded images;
4. I2V-14B-720P - a version for creating high-quality video in 720p resolution from the original images with 14 billion parameters.

All four models are available on Alibaba Cloud's ModelScope and HuggingFace platforms, and the source code is hosted on GitHub. This greatly simplifies developers' access to the technology and allows it to be integrated into various projects without the need for self-configuration from scratch.^[1]

Notes