Alibaba's Wan 3.0 video model arrives on Vercel AI Gateway with 30-second clips and audio

Alibaba's Wan 3.0 video model is now available on Vercel's AI Gateway under the model ID alibaba/wan-v3.0-video. According to the Vercel blog, a single model now covers text-to-video, image-to-video, first and last frame conditioning, and reference-based generation, and it accepts image, video, and audio as references. Clips run up to 30 seconds at 30fps in 480p, 720p, or 1080p, and the output includes audio by default.

One model where Wan 2.7 needed two

Vercel frames the release as a consolidation. The previous generation, Wan 2.7, shipped as two separate model IDs — one for text-to-video and one for reference-based video — each limited to 15 seconds at 24fps. Wan 3.0 merges those capabilities behind a single ID, doubles the maximum clip length, raises the frame rate to 30fps, and adds a 1080p option, while making audio a default part of the output rather than an extra step.

Asynchronous generation, three ways

Alongside the model, the announcement highlights asynchronous generation on AI Gateway, which executes jobs in the background rather than keeping an HTTP request open for the whole run. The blog lays out three routes, which differ in what waits:

• Polling. Passing a poll option has the gateway track the generation while the SDK issues short status requests until the video is ready. Vercel describes this as the closest option to a plain synchronous call. The published example generates a 10-second 1080p clip, checked every five seconds with a ten-minute timeout.
• Start and check later. experimental_startVideo returns as soon as the gateway accepts the job, so nothing waits and the job keeps running even after the process that submitted it exits. The returned operation handle can be fed to getVideoStatus later, from the same process or a different one. Vercel points to serverless functions, queues, and batch fan-out as the natural fits.
• Webhooks. Passing a webhook URL means the gateway posts a single event to the developer's endpoint when the generation finishes, eliminating polling entirely. The documentation also covers how to verify that a delivery is genuine.

How references work

References are supplied under inputReferences, each with its own URL and media type, alongside a prompt describing the scene. The constraints vary by kind: audio references must be hosted URLs, whereas images can also be sent as base64 data. First and last frame conditioning accepts exactly one image per boundary frame and cannot be combined with the reference list. For developers who want to evaluate the model before writing any code, it is also available in the model playground.

Why it matters

Two things are happening at once here. On the model side, Wan 3.0 collapses what used to be a multi-model workflow — choosing between a text-to-video or reference-based variant, working around a 15-second cap at 24fps, and handling audio separately — into a single API call behind one model ID. That meaningfully lowers the integration cost for teams adding video output to a product.

On the platform side, the asynchronous routes matter because a 30-second 1080p clip is not something you want to generate while holding a connection open. Background jobs surfaced through polling, operation handles, or webhooks align with the serverless and queue-driven architectures that run on Vercel, where long-lived requests are awkward or unworkable. Taken together, the release positions AI Gateway as a plausible home for production video generation rather than one-off experiments, and it puts Vercel's offering alongside platforms that already expose video models as first-class APIs.