Wan 3.0 is an all-in-one video model: what the API actually does

Alibaba updated the Wan 3.0 documentation on 27 October. The model is now positioned as an all-in-one video generator: wan3.0-video and wan3.0-video-prime handle text-to-video, image-to-video, multi-modal reference generation, video editing, and video extension through a single endpoint. You do not switch model names between tasks. The router reads the input.media types and your prompt intent.

The headline numbers: up to 30 seconds per generation at 30 fps, native audio output including dialogue, BGM, and sound effects, and up to 20 multimodal references in a single call. Resolutions are 480p, 720p, and 1080p. Aspect ratios include 16:9, 4:3, 1:1, 3:4, 9:16, and an adaptive mode that matches your input.

What "all-in-one" means in practice

Previous Wan releases separated tasks across different endpoints or required you to pick the right model variant. Wan 3.0 collapses that into one call. The routing logic is straightforward:

• Text-to-video: pass only a prompt, no media. The model generates motion, dialogue, and sound from scratch.
• Image-to-video: set type to first_frame or combine first_frame and last_frame for controlled start and end points.
• Multi-modal reference: mix up to 10 images, 5 videos (total ≤15 s), 5 audio clips (total ≤15 s), plus one document or one public web link. Reference them in your prompt as "image 1", "video 1", "audio 1".
• Video editing: pass a reference_video and use edit-intent keywords in your prompt ("remove", "replace", "change style"). Unspecified regions stay intact.
• Video extension: pass a reference_video with extend-intent keywords ("extend forward", "extend backward"). Input plus output duration must stay under 30 seconds.

Reference images, videos, and audio are counted separately. You can combine all three types in one request. Documents and links are mutually exclusive with each other but can coexist with image/video/audio references. First-frame and last-frame inputs cannot be mixed with reference types in the same call.

Native audio changes the pipeline

Wan 3.0 generates dialogue, background music, and sound effects natively. Audio is on by default; set audio=false to disable it. This removes the separate TTS, lip-sync, and sound-design passes that most current workflows require. For product demos, talking-head clips, and narrative shorts, the time saved is significant.

The trade-off is control. If you need precise audio mixing or a specific voice talent, you will still want a post-production pass. But for first drafts, social content, and internal reviews, native audio gets you to a watchable result in one step.

Document and web-to-video

Wan 3.0 accepts docx, xlsx, pptx, pdf, txt, and markdown files as direct inputs, as well as public web URLs. The model parses the content and generates video from it. This is useful for turning pitch decks, reports, and documentation into visual summaries without scripting by hand.

The limitation is access: linked pages must be publicly reachable without login. Internal wikis and gated content need to be exported as files first.

Where this fits alongside Kling 4.0

We covered Kling 4.0's API launch earlier today. The two models now occupy adjacent niches:

• Kling 4.0 leads on keyframe control (up to 10), voice-reference lip-sync via Omni, and established credit-based billing that teams already budget for. It is the stronger choice when you need precise camera choreography or multi-character dialogue with synced voices.
• Wan 3.0 leads on input flexibility (20 multimodal references, documents, web links), native all-in-one audio, and a single-endpoint workflow that reduces integration complexity. It is the stronger choice when you are working from mixed assets or want to skip the post-audio pass.

Both cap at 30 seconds. Both support first-frame and last-frame control. Neither offers native 4K yet — Kling lists it as coming soon, Wan tops out at 1080p. For 4K hero shots, Veo 3.1 Standard remains the pick.

Pricing and availability

Wan 3.0 is available through Alibaba Cloud Model Studio (DashScope). Billing is per-generation and varies by resolution and duration. Check the official guide for current rates and regional availability — model support differs by region, and cross-region calls fail silently.

The DashScope Python SDK requires version 2.0.1 or later. The API uses async job submission with polling for results. Generated video URLs expire after 24 hours, so download promptly.