Alibaba publishes video AI models as open source

The Chinese company Alibaba Cloud has announced the release of open source versions of its video-generating artificial intelligence Wan2.1. There are four models in total, half with 14 and half with 1.3 billion parameters. These should be able to generate high-quality images and videos from text and image input.

Developers, researchers and companies worldwide can download the models via the AI model community Model Scope from Alibaba Cloud and on the Hugging Face platform and use them directly, the company announced. Wan2.1 tops the VBench benchmark with an overall score of 82.22 percent and is the only open source model among the top 5 video-generating AIs. Using the text-to-video model with 1.3 billion parameters, it is possible to generate a 5-second video in 480p resolution in around four minutes on a standard notebook.

Wan2.1 is the first video generation model that supports text effects in both Chinese and English. It is also considered particularly advanced in "multi-object interaction" – the AI can create videos in which multiple objects or people interact realistically with each other, which is considered technically very demanding.

Is Alibaba shaking up the video AI market?

The release by Alibaba follows the example of the text-generating AI DeepSeek, which also comes from China. Its open source release and performance has caused quite a stir, especially in the USA. The current video AI release is also likely to further fuel the competition with US AI developers. Wan2.1 is competing in an environment where closed and expensive models have dominated to date.

Alibaba is also planning to invest around 52 billion US dollars. These are to flow into cloud computing and AI infrastructure. The Chinese technology company is known for its first open source model Qwen (Qwen-7B). Alibaba is also Apple's local partner in China to enable the introduction of Apple Intelligence there.

(mki)