Alibaba's Wan 3.0: 30-second AI video from text, documents, or a URL

Alibaba just opened the public beta of Wan 3.0, its next-generation video generation model — and the headline capability is a first for the industry: generate up to 30 seconds of video in a single pass, straight from documents, spreadsheets, and even live web pages.

Here's what's confirmed, what it costs, and what's still missing.

What Wan 3.0 does

Wan 3.0 is marketed as an "all-in-one reference video generation model" on Alibaba Cloud Model Studio (model ID wan3.0-video). One model, multiple input workflows:

• Text-to-video — a written scene and motion description become a clip
• Image-to-video — animate an image while keeping it as the visual starting point
• First-and-last-frame — anchor both ends of a shot to control a planned transition
• Reference-to-video — guide the output with reference material
• Document-to-video — the new one: feed it a PDF, PPT, DOC, XLS, or Markdown file (up to 100 MB or 50 pages) and get a video
• Web page-to-video — a live URL as input

The 30-second single-take limit is the other big change. Wan 2.7 topped out at 15 seconds per generation, forcing creators to stitch clips together. A 30-second pass means continuous camera movement and one-take shot language without visible seams.

API details and pricing

The preview API reference confirms 480P, 720P, and 1080P output, generation lengths from 2 to 30 seconds, and Beijing and Singapore endpoints. A single request can include up to 10 images, 5 videos, and 5 audio files as references — though when video is used as input, input plus output duration must stay under 30 seconds. Audio is generated by default, with seed and watermark controls exposed.

Pricing is per second of output:

• 480P — ¥0.30/s (Beijing) / ¥0.37/s (Singapore)
• 720P — ¥0.60/s / ¥0.75/s
• 1080P — ¥1.20/s / ¥1.50/s

A 30-second 1080P clip runs about ¥36 (~$5, per verified regional pricing). At 720P, roughly half that.

What's not there yet

Three caveats worth knowing before building a workflow on this:

• Access is restricted. The model is labeled preview on the API page, and the pricing page says access is invitation-only. Account and regional availability need to be verified per user.
• No confirmed 4K. The documented API ceiling is 1080P, despite third-party sites marketing "4K native".
• No open weights so far. Unlike Wan 2.7 (which shipped select models under Apache 2.0 after months of community pressure), no Wan 3.0 weights have been published — an official repository appeared on August 11 under Apache 2.0, but it currently ships only a README, with no weights or inference code. The pattern across the series — Wan 2.5 and 2.6 were API-only, 2.1 and 2.2 fully open — suggests a baseline release may eventually arrive, but nothing is confirmed.

Who should care

The document-to-video mode is the most interesting for content teams: it's the closest thing yet to "paste your deck, get a video". The 30-second ceiling matters too — it's long enough for a complete transition or short scene without stitching.

For AI video tinkerers, the usual evaluation rules apply: launch reels show the intended quality bar, not the average success rate. Test short clips (5–8s) for face and motion stability, then longer ones (20–30s) to see where identity drift and timing errors accumulate.

Sources: Alibaba Cloud Community: Wan3.0 announcement, Wan3.0 API reference, Ani Kuku spec verification.