Alibaba Wan3.0: 30-Second Video and Document Inputs

Twice the runtime, and a document as a starting point

Alibaba Cloud formally released Wan3.0 on August 24, 2026, two weeks after an application-only public beta that opened August 6. The model produces clips up to 30 seconds in a single run, twice the 15-second ceiling of its predecessor Wan2.7, at up to 1080p and 30 frames per second. [1]

A generation request can draw on text, images, audio, video and a public webpage, and it can also take a single attached document up to 100 MB, including PDF, Word, PowerPoint, Excel, plain text, Markdown and Apple's Keynote, Pages and Numbers formats, with paged formats capped at 50 pages. The document and web-link options are mutually exclusive within one request. [2]

Atlas interpretation: That document path is the functional change, not the longer runtime by itself. It lets a slide deck or spreadsheet become the source material for a video job directly, rather than requiring someone to first condense it into a text prompt. Dataconomy frames the resulting appeal as being to marketing departments and corporate communications teams as much as to filmmakers, which fits a model whose inputs now include the file formats those teams already produce work in. [1]

A reversal on open weights

Wan3.0 is not open. As of August 25, developers could reach it only through Alibaba's hosted Model Studio service, with the canonical Wan organizations on GitHub and Hugging Face carrying no Wan3.0 model files to self-host. That breaks with the prior lineage: Wan2.2 shipped as an open-source update in mid-2025, and the Wan series had built its reputation on that pattern. [2][1]

Atlas interpretation: Open weights had been part of how Alibaba competed on AI video, giving developers a model they could run and fine-tune themselves instead of renting by the second. Dataconomy reports the closed release coincides with a shift toward monetizing the capability directly: Wan3.0 is billed per generated second, at $0.05, $0.10 and $0.20 for 480p, 720p and 1080p respectively, meaning a full 30-second 1080p clip lists at $6 versus approximately $3 for the old 15-second ceiling at the same rate. [1][2]

Alibaba pitches that pricing as competitive against rivals: Dataconomy cites Google's Veo 3.1 at $0.40 per second for standard service, well above Wan3.0's top rate. Neither outlet found an independent benchmark of Wan3.0's output quality at release; a hosting company's own comparison reported stronger identity retention than Wan2.7 in some tests but also repeated unwanted cuts during shots meant to stay continuous, a test with a commercial interest in the models it sells access to. [1][2]

A day after the money for AI infrastructure was priced

Wan3.0's formal release followed by one day Alibaba's pricing of a roughly $10.2 billion Hong Kong share placement, with the company stating that all net proceeds would fund full-stack AI, including infrastructure, research and models, and AI-native applications. [2]

Atlas interpretation: The two events sit next to each other on the calendar but are not tied by any specific allocation. Winbuzzer notes the placement and Alibaba's June-quarter results, which showed AI Cloud and Compute Services revenue up 45 percent year over year against a 75 percent rise in capital spending, explain the company's investment setting rather than Wan3.0's capabilities or performance. The share sale funds the infrastructure Alibaba's cloud AI business runs on; it does not by itself explain why this particular model shipped closed rather than open. [2]