Alibaba launches Wan3.0, its 30-second video model, days after raising $10bn

Alibaba has launched Wan3.0, the latest version of its video generation model, days after raising roughly $10.2bn in a Hong Kong share placement whose proceeds it has earmarked entirely for AI. The sequencing is not subtle, and it is not meant to be, since the company has spent the past fortnight arguing to investors that its capital expenditure is buying something they can see.

The model had already been running in public beta since early August through Alibaba Cloud’s Model Studio and its Qwen Cloud platform, so Monday’s launch is a widening of access rather than a first appearance. It also lands in a category where Chinese labs have quietly built a commanding position while Western attention has stayed on chatbots.

Wan3.0 generates clips of up to 30 seconds in a single pass, double the 15-second ceiling of its predecessor Wan2.7, at resolutions up to 1080p. Alibaba says the model holds character detail, props, spatial layout, and motion graphics steady across the full length and renders faces with synchronised micro-expressions and multilingual voice output.

The more distinctive feature is what it will accept as input. Alongside text, images, audio, and video, Wan3.0 takes web pages and documents, including PDFs and PowerPoint files, which lets a user hand it a slide deck or a spreadsheet and get back a video that reflects the structure of the source material.

That points the product at marketing departments and corporate communications teams rather than at filmmakers, though Alibaba lists short dramas and social content among the use cases too, along with simulation training for autonomous vehicles and robotics. It is a business tool wearing a creative tool’s clothes.

Pricing follows the same logic. Wan3.0 runs at $0.05 per second at 480p, $0.10 at 720p, and $0.20 at 1080p, which works out at $12 a minute for the top tier against $0.40 per second for the standard tier of Google’s Veo 3.1.

Undercutting the incumbent is a familiar move, though the arithmetic is not entirely one-sided. Several rivals that currently rank above Alibaba’s previous model on independent leaderboards also cost less per minute than Wan3.0 does at 1080p, and Wan3.0 itself has not yet been independently benchmarked at all, which makes the quality claims company-issued for now.

Price pressure in the category is coming from further down the market as well, with an Indian startup recently shipping a video model at half a cent per second and betting that most commercial video never needs frontier quality.

Access is still staged rather than open. The beta runs through Model Studio and Qwen Cloud on an application basis, full API availability is described as rolling out, and a consumer-facing site at wan.video is promised as a members-only platform, which suggests Alibaba is metering demand while it works out what the compute costs it to serve.

Notably, the weights are closed. Alibaba built much of its reputation in this field by open-sourcing the Wan series, and Wan2.2 in July 2025 remains the last video flagship it released that way, a shift that sits alongside its decision to start charging the heaviest users of its open Qwen models.

The commercial case for all of this showed up in the last set of results. Cloud and AI revenue grew 45% to 48.44bn yuan in the June quarter, and AI model services alone now carry more than 16bn yuan in annualised recurring revenue, which is the line the share sale is meant to keep growing.

The cost of getting there was equally visible, with quarterly net profit falling 75% as capital expenditure rose 75% to 67.68bn yuan. Alibaba committed 380bn yuan to AI infrastructure over three years in early 2025 and has been reported to be weighing an increase to 480bn.

Whether Wan3.0 justifies any of that is not a question this week can answer. Video models are expensive to run, buyers switch based on price, and the leaderboards that would settle the argument have not yet had a look at it.