Wan 3 video

Wan 3 is Alibaba Cloud Model Studio's unified premium video model. On Sogni, every supported mode uses the single model ID wan3.0-video: text generation, first-frame and first-and-last-frame animation, loose image/video/audio references, audio-driven generation, and document or webpage context.

Alibaba currently labels the upstream Wan 3 API as preview (as of August 2026).

Wan 3 renders 2–30 second MP4 clips at a fixed 30 fps in 480p, 720p, or 1080p. Native audio is enabled by default, provider prompt expansion is enabled by default, and both can be controlled per request.

#Capabilities

Wan 3 video inputs are loose references for a new generation. Alibaba's Wan 3 API does not expose a source-video edit task, an extend task, or a task-type parameter. A prompt can ask for a transformation or continuation inspired by Video 1, but the provider does not promise frame-level preservation or append-only extension. Use an editing-capable model or Sogni's composition tools when unchanged source footage must be preserved.

#Generation settings

• Resolution: 480, 720, or 1080 as the short-side target.
• Ratio: adaptive, 16:9, 4:3, 1:1, 3:4, or 9:16.
• Duration: 2–30 seconds, sent as an explicit duration. When video references are present, their total input duration plus the requested output duration cannot exceed 30 seconds. (smartDuration, which let Wan 3 pick the length, is retired: it forced the quote to reserve the 30-second maximum and billed against that ceiling rather than the delivered video.)
• Frame rate: fixed 30 fps.
• Audio: on by default. Set generateAudio: false only for an explicitly silent output. Audio on and off cost the same.
• Watermark: off by default. Set watermark: true to add Alibaba's visible watermark.
• Seed: an integer from 0 through 2,147,483,647.
• Prompt: up to 20,000 characters. Wan 3 has no negative-prompt, steps, guidance, sampler, scheduler, ControlNet, or mask request fields.

With smart duration, Sogni reserves the maximum 30-second cost before starting. Once Alibaba reports the completed output duration, Sogni settles the job down to its actual cost.

#Reference inputs

A loose-reference request accepts up to:

• 10 images
• 5 videos, each 1–15 seconds and no more than 15 seconds total
• 5 audio clips, each 1–15 seconds and no more than 15 seconds total

There is no additional aggregate media-count cap beyond those per-type limits. Address loose references as Image 1, Video 1, and Audio 1 in each media type's submission order, and give every reference an explicit job in the prompt.

Dedicated first/last-frame anchors cannot be mixed with loose image, video, or audio references.

#Document and webpage context

Wan 3 also accepts either:

• one public HTTPS document URL through referenceFileUrl; or
• one public HTTPS webpage URL through referenceLinkUrl.

The two are mutually exclusive. A document can be DOCX, DOC, XLSX, XLS, PPTX, PPT, PDF, TXT, Keynote, Pages, Numbers, or Markdown, up to 100 MB. PDF, DOCX, DOC, PPTX, PPT, Keynote, and Pages inputs are limited to 50 pages. Document/web context may accompany loose-reference generation, but it cannot be combined with dedicated first/last-frame anchors.

#Prompt expansion

There are two possible prompt-processing layers, and Sogni prevents them from rewriting the prompt twice:

1. Direct SDK or CLI request: Alibaba's prompt_extend behavior defaults on. Set promptExtend: false in the SDK, or --no-expand-prompt in the Creative Agent CLI, when wording must remain exact.
2. Sogni Chat or Creative Agent tool: expandPrompt controls Sogni's video prompt shaper. After Sogni successfully shapes the brief, the handler disables Alibaba's second expansion. If the shaper is unavailable, Alibaba expansion remains available as the fallback.

For an explicit "use this prompt exactly" request in a hosted tool call, set both skipPromptProcessing: true and expandPrompt: false. Ordinary creative requests should keep expansion enabled.

Strong Wan 3 prompts use plain natural language. Describe the shot, subject, environment, action beats, camera, dialogue, and sound. Quote exact spoken words and identify the speaker. For references, name each input exactly and state what must be borrowed or preserved.

#Pricing

Wan 3 is priced per second of finished video, using the standard 1 Spark = $0.005 render-value peg. Alibaba is running a 30%-off promotion and Sogni passes it through, so the rates below are the promotional ones, verified against the live estimate on 3 September 2026. Alibaba schedules the promotion through 23 September 2026 at 16:00 UTC, after which the list rate applies.

List rates, for reference once the promotion ends, are 13 / 26 / 52 Spark ($0.065 / $0.13 / $0.26) per second at 480p / 720p / 1080p.

Duration and resolution are the price factors. Workflow, aspect ratio, reference count, prompt expansion, watermark, and native audio do not change the rate. Smart-duration jobs are settled against the actual completed output duration.

Wan 3 is a Premium Spark-only external vendor model. It is not covered by an Unlimited subscription, although an active Unlimited plan's normal partner-model discount can apply. Use the job estimate calculator for a live quote.

#Sogni Web

Sogni Web exposes dedicated text-to-video, first/last-frame, and multimodal reference surfaces. The controls expose fixed or smart duration, every supported ratio and resolution, audio, prompt expansion, watermark, and document/web context where the upstream contract permits it.

#JavaScript SDK examples

Fixed-duration text-to-video:

const project = await sogni.projects.create({
  type: 'video',
  network: 'fast',
  modelId: 'wan3.0-video',
  positivePrompt: 'A documentary presenter walks through a detailed workshop and says "Welcome to the future" as the camera tracks beside her.',
  duration: 8,
  width: 1920,
  height: 1080,
  ratio: '16:9',
  generateAudio: true,
  promptExtend: true,
  watermark: false,
  tokenType: 'spark'
});

const urls = await project.waitForCompletion();

Smart-duration document-guided generation:

const project = await sogni.projects.create({
  type: 'video',
  network: 'fast',
  modelId: 'wan3.0-video',
  positivePrompt: 'Present the supplied launch brief as a concise cinematic product story.',
  duration: 10,
  width: 720,
  height: 1280,
  ratio: '9:16',
  referenceFileUrl: 'https://cdn.example.com/launch-brief.pdf',
  promptExtend: false,
  tokenType: 'spark'
});

For video-guided generation, attach clips through the SDK's loose video-reference fields and refer to them as Video 1, Video 2, and so on. Do not send wan3TaskType; it is a deprecated Sogni compatibility field and has no Alibaba API equivalent.

#Creative Agent Skill examples

# Direct Wan 3 generation with provider expansion on
sogni-agent --video -m wan3 --target-resolution 1080 --duration 8 'A presenter says "Welcome." in a detailed studio'

# First and last frame
sogni-agent --video -m wan3 --ref first.png --ref-end last.png 'Move smoothly between the supplied frames'

# Multimodal references
sogni-agent --video -m wan3 --workflow ia2v --ref presenter.png --ref-audio dialogue.mp3 'Use Image 1 and Audio 1 for the performance'

# Smart timing, exact prompt, and document context
sogni-agent --video -m wan3 --smart-duration --wan3-ratio 9:16 --no-expand-prompt --reference-file-url https://cdn.example.com/brief.pdf 'Use the supplied brief exactly'

# Loose video-reference generation (best-effort conditioning, not append-only extension)
sogni-agent --video -m wan3 --workflow r2v --ref-video source.mp4 'Use Video 1 as motion and style inspiration for a new continuation shot'

Official sources: Alibaba Wan 3 API reference and Wan 3 model release and base provider pricing.

Need help? Ask in our Discord!

For questions about the model lineup, requests for new models, or tips — join the Sogni Discord ✨