Alibaba: wan-3.0

Wan 3.0 is Alibaba Tongyi Lab's video generation model, released as a public beta on August 6, 2026. Built on the diffusion-transformer paradigm refined across the Wan series, it generates native video of up to 30 seconds in a single continuous take and produces synchronized audio in the same pass. Omni-Reference lets the model accept images, video clips, audio, documents, spreadsheets, slides, PDFs, and webpage URLs, with up to 20 reference inputs per generation that can be addressed positionally in the prompt. Output supports every aspect ratio from 16:9 to 9:16 and three resolution tiers: 480p, 720p, and 1080p. It is available through Alibaba Cloud Model Studio and Qwen Cloud under the model ID wan3.0-video, and on wan.video and third-party platforms. The model targets AI film and short-form storytelling, corporate communications and quarterly business updates, marketing explainers, short-form social content for TikTok, Instagram Reels and YouTube Shorts, e-commerce and advertising product shots, and visual restyling or scene changes on existing video. Strengths include strong visual fidelity, motion coherence, and prompt adherence, with reliable rendering of specified lighting, object placement, and action sequences. Scene consistency stays stable across the full clip length in static or slow-moving shots, and surface-detail retention holds label text and specular highlights for product shots. Document-to-video and webpage-to-video are described as an industry first. The model is closed-weight with no local deployment option and is currently in API preview with access approval required. It requires substantial GPU resources for inference, with estimates suggesting 4 to 8 A100 or H100 GPUs for the full checkpoint. Dynamic scenes with multiple interacting subjects can show drift in secondary character positioning, and independent benchmarks are still limited. The public beta does not support 4K output.