Wan 3.0

Launched in beta in August 2026, Wan 3.0 is Alibaba's latest video generation model. It produces videos up to 30 seconds long — doubling Wan 2.5's length — and accepts an unusually broad set of inputs: text prompts, up to ten images, five videos, five audio clips, plus web pages and documents (PDFs, PowerPoint) that it turns into dynamic video. A single prompt can mix modalities, and the model recommends the best video length for the prompt plus an extension tool to lengthen existing clips. It targets visual drift and distortion (faces, UIs) with better consistency of characters, props, and spatial layouts. Available via wan.video, Alibaba Cloud Model Studio, and Qwen Cloud API, with Standard and faster Prime tiers (e.g. $0.20/sec Standard at 1080p).

Limitations: in beta with paid, metered API access — and unlike Wan 2.1/2.2, open weights are unverified. Widely repeated claims of Apache 2.0 weights on Hugging Face are unsupported by Alibaba's own announcements; treat the model as closed until a license and repo appear. Beta pricing and availability may change.

No public benchmark numbers located; vendor claims on consistency