Wan 3.0 Now Available on AI Gateway

Alibaba's Wan 3.0 video model has been released on Vercel AI Gateway under the ID alibaba/wan-v3.0-video. Unlike previous versions, it integrates text-to-video, image-to-video, first/last frame conditional generation, and reference-based generation into a single model.

Wan 3.0 generates videos up to 30 seconds long at 30fps with resolutions from 480p to 1080p, and supports audio output by default. This contrasts with the previous version, Wan 2.7, which had a 15-second/24fps limit and used separate model IDs.

AI Gateway supports asynchronous generation, processing requests in the background. It provides three methods for this:

• poll: The SDK periodically checks the status, representing the simplest change.
• experimental_startVideo: Returns immediately upon job acceptance, making it suitable for serverless functions or batch processing.
• webhook: Sends an event to a specified endpoint upon generation completion, eliminating polling entirely.

Reference materials can be passed via inputReferences by specifying URLs and media types. Audio references only support hosted URLs, while images also allow base64. First/last frame conditional generation uses one image each and cannot be used together with reference materials.