Wan 3.0

Wan 3.0 is Alibaba Tongyi Lab's latest video model, in public beta since 6 August 2026. It generates native 30-second clips at up to 1080p with audio in the same pass. Its signature feature is Omni-Reference: alongside text, image, audio, and video, it accepts documents, spreadsheets, slides, pdfs, and webpages and builds video from them. It is available through Alibaba Cloud Model Studio and Qwen Cloud as the model wan3.0-video.

Yes. Alibaba launched Wan 3.0 in public beta on 6 August 2026, on Alibaba Cloud Model Studio and Qwen Cloud, with wider access on wan.video rolling out to members. The model ID (wan3.0-video), the resolution tiers, and per-second API pricing are all now published, so it is a live model rather than the rumor it was through July 2026.

Up to 30 seconds in a single run, with intelligent duration that matches clip length to the prompt so short scenes are not padded. That is double the 15-second ceiling on Wan 2.7, and it is built for one continuous take rather than short fragments stitched together.

No. Wan 3.0 is a closed, API-only model, and no weights have been published on Hugging Face, GitHub, or ModelScope. Alibaba's last open-weight video flagship remains Wan 2.2 under Apache 2.0, released in 2025. If self-hosting is a requirement, Wan 2.2 is still the newest downloadable Wan, and any site advertising Wan 3.0 weights or a free download is not offering the real model.

Wan 3.0 accepts doc, xls, ppt, pdf, txt, key, pages, numbers, and md files, plus webpage URLs. It reads the content and generates a video sequence from it, which Alibaba describes as turning static, text-heavy data into finished video. This document- and webpage-to-video capability is the main thing that sets Wan 3.0 apart from other current video models.

Both launched within days of each other in early August 2026, and both generate 30-second single-pass clips with audio. Seedance 2.5 accepts more reference inputs, up to 50 across images, video, and audio, while Wan 3.0's distinguishing capability is building video from documents and webpages. Neither model has independent benchmarks yet, and both are closed, API-only releases, so any head-to-head today rests on demos rather than measured scores.