Skip to main content
Back

Alibaba's Wan 3.0 Generates 30-Second Video and Audio Through one API

Alibaba's new all-in-one video model tops the Artificial Analysis editing leaderboard, generates 30 second 1080p clips with native audio, and accepts documents as input.

  • Wan 3.0 debuts at #1 on the Artificial Analysis Video Editing leaderboard and #2 in Text to Video with Audio.
  • Generates up to 30 seconds of 1080p video at 30fps in a single pass, with native dialogue, music, and sound effects.
  • Accepts text, images, audio, video, and for the first time PDF, DOC, XLS, PPT, and web page inputs.
  • Priced at $0.05, $0.10, and $0.20 per second for 480p, 720p, and 1080p on Alibaba Cloud Model Studio.
  • Closed weights and hosted only, breaking from the open-source lineage of Wan 2.2.
  • Known weak spots include on-screen text rendering, audio texture, and image-to-video quality relative to Alibaba's own HappyHorse line.

Alibaba’s Tongyi Lab has released Wan 3.0, a hosted audiovisual model that handles text-to-video, image-to-video, reference-guided generation, editing, and clip extension. It produces synchronized dialogue, music, and sound effects, supports document inputs, and can generate up to 30 seconds of video at 30 frames per second.

Combining those workflows can reduce the number of models and tools required to produce a clip. A team can supply source material, generate video and audio, then revise visuals, plot, or dialogue through the same service. Developers still need to account for regional availability, input limits, uneven text rendering, and the absence of downloadable weights.

Wan climbs the arena

The Video Arena ranks models through side-by-side user preferences. At publication, Wan 3.0 had improved on Wan 2.7 across all three audiovisual categories, although leaderboard positions can change as votes and models are added.

Artificial Analysis Video Arena rankings at publication
Category Wan 3.0 Wan 2.7
Video editing with audio No. 1 No. 5
Text-to-video with audio No. 2 No. 6
Image-to-video with audio No. 5 No. 12

The results provide human-preference data across broad categories. Teams evaluating production use will still need task-specific tests for prompt adherence, temporal consistency, latency, failure rates, and audio quality.

One request, five production paths

Wan 3.0 supports text-to-video, image-to-video, reference-to-video, video editing, and extension. A single generation can run for up to 30 seconds at 30 fps, doubling Wan 2.7’s 15-second limit. Extension mode can continue a clip beyond the initial generation.

The model generates dialogue, background music, and sound effects alongside the visuals. Requests can include reference audio or a first-frame image, while outputs support 16:9, 4:3, 1:1, 3:4, and 9:16 aspect ratios.

Alibaba says the model improves realistic face rendering and consistency between reference material and generated footage. Those capabilities become harder to sustain as clips lengthen because small changes in faces, objects, lighting, and camera position can accumulate across frames.

Documents become source material

Wan 3.0 accepts DOC, XLS, PPT, PDF, and Markdown files in addition to text, images, audio, and video. Each request can include one document or link, with limits of 100 MB and 50 pages.

That input layer allows a slide deck to supply product claims and visual references for an advertisement, a training document to guide courseware, or a spreadsheet to drive animated charts. Alibaba presents these jobs as one-call workflows, reducing handoffs among scripting, voice, video, and editing tools.

Document support treats the file as source material for generation. Teams that require exact wording, chart values, legal copy, or branded typography should validate the output because the model’s on-screen text rendering remains inconsistent.

Closed weights, metered output

Alibaba distributes Wan 3.0 exclusively through Model Studio, and downloadable model weights are unavailable. Wan 2.2, released in July 2025, remains the newest open-weight video release in the Wan line.

Wan 3.0 output pricing
Resolution List price per second List price for 30 seconds Promotional price per second Promotional price for 30 seconds
480p $0.05 $1.50 $0.035 $1.05
720p $0.10 $3.00 $0.07 $2.10
1080p $0.20 $6.00 $0.14 $4.20

Alibaba lists the promotional rates through September 24, 2026. Google’s Veo 3.1 Standard service costs $0.40 per output second, providing a rough list-price comparison. Production costs will also depend on retries, discarded generations, storage, and any surrounding processing pipeline.

Thirty seconds exposes drift

Longer clips give temporal errors more time to surface, including changing faces, displaced objects, unstable backgrounds, and inconsistent motion. Alibaba says Wan 3.0 uses higher-precision visual continuity to reduce that drift and includes an intelligent-duration feature that recommends a clip length from the prompt.

The editing workflow can modify visuals, plot, or dialogue in an existing clip. Targeted editing may save time when a team needs to revise one scene or spoken line, although the resulting continuity and audio synchronization require testing on each production pattern.

Limits that affect production

Alibaba’s materials identify audio texture and on-screen text accuracy as areas still under development. The text limitation directly affects lower thirds, title cards, interfaces, packaging, and product shots containing visible copy.

  • Reference video accepts up to five files with a combined duration of 15 seconds.
  • Reference audio accepts up to five files with a combined duration of 15 seconds.
  • Document and link input is capped at one item, 100 MB, and 50 pages per request.
  • Model availability varies by region. The model deployment, API key, and endpoint must use the same region.
  • Official output tiers stop at 1080p. Alibaba’s pricing and model documentation list no 4K option.

Model Studio documentation recommends Alibaba’s HappyHorse line as the default for image-to-video and reference-to-video jobs, directing users to Wan when custom audio upload is required. That guidance, combined with Wan 3.0’s No. 5 image-to-video ranking, gives developers a reason to benchmark both product lines before standardizing an integration.

From campaign clips to simulation

Alibaba released Wan 3.0 days after raising roughly $10.2 billion through a Hong Kong share placement whose proceeds were earmarked for artificial intelligence. The launch supplies a commercial product for that investment strategy, with usage-based pricing and a closed hosted service.

Alibaba also connects richer multimodal inputs and more consistent video generation to longer-term work in embodied AI, autonomous driving, and industrial simulation. Applying the technology in those fields will require stronger evidence of physical accuracy, controllability, repeatability, and safety. Wan 3.0 currently broadens the inputs and workflows available through a production video API while exposing the technical limits developers must design around.