What is Wan 3.0?
Wan 3.0 is Alibaba's video generation model, in public beta since August 6, 2026. It is not yet callable on Unifically. We will open it here once the integration is live, and every spec on this page comes from the beta announcement and will be re-verified against the callable API on day one.
The model generates up to 30 seconds of video in a single pass, double the 15-second ceiling of Wan 2.7. The bigger change is what it takes as input: text, images, video, and audio, plus webpages, PDFs, and PowerPoint files, with up to 20 reference assets in one generation. Upload a product deck, add a prompt, and the output is a finished promotional video.
Audio generates with the picture: multilingual voices, sound design, and stable audio carried over from reference inputs.
Key features of Wan 3.0
- Up to 30 seconds in one generation. Room for complete camera moves and unbroken shots instead of joining short videos. An intelligent duration feature picks the length that fits the prompt, and extension features grow a finished video further.
- Documents and webpages as input. PDFs, PowerPoint files, and webpages parse directly into video, alongside text, images, video, and audio.
- Up to 20 reference assets. One generation can carry a full set of product shots, style frames, and audio cues.
- Pixel-accurate reference consistency. Characters, props, spatial layouts, styles, and audio reproduce from references with strict detail control, not rough resemblance.
- Faces and interfaces that hold. Realistic human faces with synchronized micro-expressions, plus stable software interfaces and motion graphics, two common failure points for video models.
- Precision video editing. Instruction-based edits from a text description, or reference-based edits guided by an asset.
Best for
Long single-take stories
A 30-second ceiling covers a full narrative arc, with camera moves that develop instead of resetting.
Slides and documents to video
A product PPT or PDF plus one prompt returns a finished marketing or training video.
Reference-heavy production
Up to 20 assets in one request, reproduced with strict character and product detail.
Dialogue and sound-led scenes
Multilingual voices and sound design generate with the picture in the same pass.
Editing finished videos
Describe a change in text or guide it with a reference asset instead of regenerating from scratch.
Use cases
Turn a product deck into a launch video: the slides carry the facts, the prompt sets the tone, and the model returns one finished piece with voice-over. Build 30-second ads as one continuous shot instead of stitching three generations. Convert documentation and reports into training videos for teams that will not read a PDF. For short dramas and social stories, the longer timeline plus micro-expression rendering keeps a performance running through multiple beats. Alibaba also positions it for simulation video that trains self-driving and robotics systems, where stable physics across a long take matters.
Limitations
Access is beta-only as of August 17, 2026: you apply for testing through Alibaba Cloud's Model Studio or Qwen Cloud, and there is no public API pricing. Alibaba has not published output resolution or frame rate specs; the demo videos released with the announcement are 720p. Open weights are not announced for this release, unlike the Wan 2.1 and 2.2 open models. And Arena lists no Wan 3.0 entry on its video boards yet, so there is no independent rank or Elo for it, only Alibaba's own material.
Wan 3.0 vs Wan 2.7
Wan 3.0 doubles the single-pass ceiling to 30 seconds against Wan 2.7's 15, and adds input types 2.7 does not have: webpages, PDFs, PowerPoint files, and a 20-asset reference budget. Both generations cover editing, but 3.0 folds instruction-based and reference-based editing into the main model where 2.7 splits it into a separate edit route. The practical difference today is availability. Wan 2.7 is live on Unifically now with text-to-video, image-to-video, and reference-to-video at 720P or 1080P, while Wan 3.0 is an application-only beta. Use Wan 2.7 for work due this month; move the long-form and document-driven jobs to 3.0 when it lands here.
Availability
Alibaba announced the Wan 3.0 public beta on August 6, 2026. Testing runs through Alibaba Cloud's Model Studio and Qwen Cloud, with the model also surfaced across Alibaba's own creation apps. There is no third-party API access or published pricing yet. Unifically pricing, variants, and exact callable parameters will appear on this page once the model is live here.
What Wan 3.0 can do
30 seconds in a single generation
One continuous 30-second action sequence: combat robots move through a ruined city while the camera tracks, cuts, and reframes without the scene resetting. This is the length that used to take three or four separate generations.
One model, 20 reference assets
Alibaba's Omni-Creation demo runs a 30-second desert story from a wide gas-station establishing shot to a close-up performance, with faces, wardrobe, and setting holding steady across every cut. Reference inputs, including documents and webpages, feed one generation.
Instruction-based video editing
The Precision Video Editing demo rebuilds a refrigerator interior into a miniature world, with a character diving past egg cartons and sledding through broccoli forests. Edits follow a text instruction or a reference asset instead of a full regeneration.
Reference details that survive the render
A violinist plays through heavy rain with wet hair, soaked fabric, and instrument detail staying coherent under hard motion and low light. The Pixel-Perfect Consistency demo shows the strict reference control Alibaba is claiming for characters and props.
Sound design generated with the picture
A stylized warrior draws a glowing blade as water beads scatter, with the audio rendered in the same pass as the frames it belongs to. Voices come back multilingual, and audio from reference inputs stays stable through the generation.
FAQs
People also ask
Wan 3.0 is Alibaba's video generation model, in public beta since August 6, 2026. It generates up to 30 seconds of video in a single pass and accepts text, images, video, audio, webpages, PDFs, and PowerPoint files as input, with up to 20 reference assets in one generation.
Not yet. Alibaba opened a public beta on August 6, 2026, with testing access through Alibaba Cloud's Model Studio and Qwen Cloud. We will open it on Unifically once the integration is live, and this page will get the playground, callable parameters, and pricing at that point.
Up to 30 seconds in one generation, double the 15-second ceiling of Wan 2.7. An intelligent duration feature can also pick the length that fits your prompt, and extension features grow a finished video further.
Text, images, video, audio, webpages, PDFs, and PowerPoint presentations, with up to 20 reference assets in a single generation. A slide deck plus a prompt is enough to produce a finished video.
Yes. The model produces multilingual voice output and sound design together with the picture, and keeps audio from reference inputs stable across the generation.
Yes. It supports instruction-based editing, where you describe the change in text, and reference-based editing, where an asset guides the change.
Alibaba has not published output resolution or frame rate specs yet. The demo videos published at the beta announcement are 720p. We will list exact callable resolutions here when the model goes live on Unifically.
There is no Wan 3.0 entry on the Arena video boards yet as of August 17, 2026. Earlier Wan releases through 2.7 are listed, so a ranking should follow once the beta output is widely testable.
