Alibaba opened the Wan 3.0 public beta, and the headline number is duration: 30 seconds generated natively in a single pass. Every long-form workflow up to now has been stitching — chain a five- or ten-second clip to the next and hope the character survives the cut. A native 30-second shot means continuous camera movement and one-take shot language become a prompt rather than an edit, which is the difference between generating footage and generating a scene. The second claim is what Alibaba calls reality-grade rendering — a push at the photographic register Kling 3.0 and Veo have been trading on, backed by production-grade character consistency held across the full 30 seconds. The third is the genuinely novel one: Omni-Reference. Wan 3.0 takes text, images, audio, and video as references like its predecessors, and adds documents, spreadsheets, slide decks, PDFs, and live web pages. Point it at a deck or a product page and it reads the source, then renders video from it — a direct shot at the marketing and explainer work that currently starts with a human writing a script from that same deck. For the open-source crowd this is the uncomfortable part. Wan built its following on downloadable weights — 2.1, VACE, and 2.2 all ran locally, and the 2.6 reference-to-video release was China's first open one of its kind. Wan 3.0 shipped API-only: reachable through Model Studio, the Wan site, and Qwen's creation tools, billed per second at roughly $0.05 / $0.10 / $0.20 for 480p / 720p / 1080p, with no weights on Hugging Face or ModelScope and no technical report, model card, or parameter count. Wan 2.2 remains the last open flagship. A formal launch event in Hangzhou on August 10 is expected to fill in the business story; for now the specs are being taken on trust.