Wan (Video Generation Model)

Wan (Chinese: 万相) is a family of generative artificial intelligence video models developed by Alibaba, the Chinese e-commerce and cloud computing group, through its Tongyi Lab and Alibaba Cloud divisions. The models generate short video clips from text descriptions or still images, and the family attracted particular attention because Alibaba released several of its largest models as open weights, allowing developers to run high-quality text-to-video generation on their own hardware rather than only through a vendor API.^[1]

History

Alibaba introduced the Wan brand as its unified video-generation family in early 2025. In February 2025 the company open-sourced Wan 2.1, releasing text-to-video and image-to-video models of roughly 1.3 and 14 billion parameters under the permissive Apache 2.0 licence, together with technical documentation. The release was significant because most comparable frontier video models — such as OpenAI's Sora or Google's Veo — were available only as closed, hosted services, whereas Wan 2.1 could be downloaded and run locally, and was quickly integrated into open-source tooling such as ComfyUI. In May 2025 Alibaba released Wan 2.2, extending the family to higher resolutions and adding smaller, faster variants able to generate 1080p video at interactive speeds on high-end consumer graphics cards. Later 2025 releases added video-editing and consistency capabilities built around the same core models.^[1][2]

Key Concepts and Technology

Wan models are text-to-video (T2V) and image-to-video (I2V) systems: a T2V model turns a natural-language prompt into a moving clip, while an I2V model animates a supplied starting image, which gives creators control over the first frame of a scene. The open-weights releases include models of different sizes, letting users trade quality against hardware requirements — smaller models target graphics cards affordable to individual creators, while the largest models aim at production quality and are typically run on server GPUs.

Alibaba distributes Wan through several routes. The model weights are published on open repositories such as GitHub and ModelScope, Alibaba's model-sharing platform, for self-hosting. Commercially, Wan is offered through Alibaba Cloud's model platform, where developers call it as an API for enterprise workloads such as advertising, e-commerce product video and short-form content. Within China, Alibaba has integrated Wan into its own consumer and e-commerce ecosystems, where automated product videos and marketing clips are common use cases. Independent benchmark evaluations, including the widely cited VBench leaderboard, placed Wan among the leading open video models of 2025, and its open licence made it a reference point in the broader debate over whether frontier generative models should be published or kept behind closed APIs.^[1][3]

Applications and Impact

Because Wan can be self-hosted, its applications differ from those of closed services. Individual creators and small studios run it locally for short-form video, concept visualisation and content prototyping, often inside free open-source interfaces rather than paid subscriptions. Companies with confidentiality concerns — film and advertising agencies handling unreleased campaigns, or enterprises that do not want creative materials sent to overseas APIs — can operate Wan on private infrastructure. Developers building commercial products use the hosted API for scale. The main practical limitations of local deployments are hardware cost and the skill required to operate GPU servers, which keeps many smaller users on hosted services despite the open weights.^[2][3]

>See Also

• Kling AI
• Sora
• Seedance
• Veo
• ComfyUI
• Open-Source AI