Last checked August 8, 2026: Wan 3.0 is real and in public beta. Alibaba Cloud Model Studio added wan3.0-video to its release log on August 6, and Alibaba Cloud published a matching public-beta demo. The official material confirms text-to-video, image-to-video, first-and-last-frame control, reference-to-video, and clips up to 30 seconds.
Alibaba Cloud has since published the preview API reference and regional prices. The documentation now confirms 480P, 720P, and 1080P output; 2-to-30-second generation; Beijing and Singapore endpoints; and reference limits. Access is still not frictionless: the API page labels the model preview, while the pricing page labels it invitation-only. Parameter count, 4K output, and downloadable Wan 3.0 weights remain unverified.
Wan 3.0 Quick Answer
| Question | Verified answer on August 8, 2026 |
|---|---|
| Is Wan 3.0 released? | Yes. It entered public beta on August 6, 2026. |
| What is the model ID? | wan3.0-video in the Alibaba Cloud Model Studio release log. |
| Which generation modes are confirmed? | Text-to-video, image-to-video, first-frame, first-and-last-frame, and reference-to-video. |
| How long can a clip be? | Up to 30 seconds. |
| Which API resolutions are documented? | 480P, 720P, and 1080P. |
| How much does the API cost? | Beijing: ¥0.30/¥0.60/¥1.20 per second; Singapore: ¥0.37471/¥0.74942/¥1.49884 per second, for 480P/720P/1080P respectively. |
| Is it generally available? | No. The API is marked preview and the pricing page says invitation-only; account and regional access still need verification. |
| Is 4K output confirmed? | No. The documented API ceiling is 1080P. |
| Is Wan 3.0 open source? | No Wan 3.0 weights, inference repository, or license were found in the official Wan organizations reviewed. |
This distinction matters because the search results mix three different things: an official model listing, access through a particular vendor or account, and claims copied by third-party Wan 3.0 sites. Only the first is universal evidence. API access, settings, and billing still need to be checked in the account that will run the workload.
What Alibaba Actually Released
Alibaba Cloud describes Wan 3.0 as an All-in-One reference video generation model. That wording is more useful than a generic quality claim: it signals that several input workflows are packaged under one model ID instead of being presented as unrelated products.
The official release entry confirms:
- Text-to-video: create a clip from a written scene and motion description.
- Image-to-video: animate an image while using it as the visual starting point.
- First-and-last-frame video: anchor both ends of a shot to control a transition or planned destination.
- Reference-to-video: use reference material to guide the generated result.
- Up to 30 seconds: produce a longer single request than the short clips common in many earlier video-model workflows.
The later API reference fills in the operational limits. wan3.0-video accepts 2-to-30-second output at 480P, 720P, or 1080P. A request can include up to 10 images, 5 videos, and 5 audio files as references. When video input is present, the total duration of the input videos plus the requested output cannot exceed 30 seconds. The API generates audio by default and also exposes seed and watermark controls.
Those limits still depend on the chosen input mode. A request that uses several video references has less output-time headroom than a text-only request. The model, endpoint, API key, and uploaded assets also need to stay in the same region.
Wan 3.0 Demo Videos
The first clip comes from Alibaba Cloud and is useful as release evidence. It is a selected launch reel, so it shows the intended quality bar rather than an average success rate.
The second video is an independent continuity test. Its value is in the questions it asks—whether subjects, scenes, and story logic survive across generated shots—not in treating one creator's result as a universal score.
When judging any Wan 3.0 demo, look for the parts that launch compilations tend to hide: failed takes, prompt and reference inputs, generation time, exact provider, model ID, region, output settings, and how many attempts were needed. A visually impressive clip without that context is inspiration, not purchasing evidence.
Why the 30-Second Limit Matters
Longer output is not automatically a better shot. It changes where problems can appear.
For motion comics and short drama, a 30-second generation can hold a complete exchange, camera move, or transition without stitching several short outputs. That may reduce edit points and visible continuity breaks. The same duration also gives identity drift, background resets, timing errors, and unwanted camera movement more time to accumulate.
A sensible evaluation uses both short and long cases:
- a 5-to-8-second close-up for face, lip, and micro-motion stability;
- a 10-to-15-second two-character shot with one clear action;
- a 20-to-30-second scene with a planned beginning, middle, and end;
- a first-and-last-frame transition where the destination frame must remain recognizable;
- a reference-to-video shot that tests costume, prop, and scene continuity.
Record the first usable take, not only the best take. If a 30-second clip takes six retries to preserve the character, its effective cost and review time can exceed a shorter workflow assembled from approved shots.
Wan 3.0 API Status
Alibaba Cloud now documents asynchronous Wan 3.0 endpoints for Beijing and Singapore. The model ID is wan3.0-video, but the API reference still marks it preview; the Chinese pricing table says the model is in invitation testing. In practice, “publicly documented” and “callable from every account” are not the same status. Confirm model visibility, activation, quota, and regional access before committing a production integration.
The integration record should include:
- region, endpoint, and account tier;
- exact model ID and request date;
- accepted text, image, first-frame, last-frame, video, audio, file, and public-URL reference fields;
- the 2–30 second duration, 480P/720P/1080P resolution, audio, seed, and watermark controls actually exposed;
- queue, concurrency, timeout, cancellation, and retry behavior;
- output URL lifetime and retention rules;
- billing unit and price shown at submission time.
Do not copy request fields from Wan 2.x and assume they still apply. Keep the provider schema behind a small adapter, preserve the raw task ID and settings, and copy approved outputs into storage you control before temporary URLs expire. For reference-heavy requests, validate the documented maximums of 10 images, 5 videos, and 5 audio files before submitting the task.
Wan 3.0 Pricing
Alibaba Cloud's current price table charges Wan 3.0 by output second and resolution. Beijing is listed at ¥0.30 for 480P, ¥0.60 for 720P, and ¥1.20 for 1080P per second. The Singapore international region is listed at ¥0.37471, ¥0.74942, and ¥1.49884 per second for the same resolutions. A 10-second request therefore has a list-price generation charge of ¥3/¥6/¥12 in Beijing or ¥3.7471/¥7.4942/¥14.9884 in Singapore, before retries and review.
These are documented list prices, not proof that an uninvited account can submit the model. Save five pieces of context with every quote: provider, region, billing unit, output settings, and date. Then budget the complete approval loop:
cost per approved second = (generation + retries + review + storage) ÷ approved final seconds
This is more revealing than “price per video.” A provider that charges less per request can cost more per finished minute if it needs extra retries for reference fidelity, dialogue timing, or physical interaction.
Is Wan 3.0 Open Source?
Not as of this review. The Wan-AI Hugging Face organization and Wan-Video GitHub organization contain official Wan projects and Wan 2.x artifacts, but no Wan 3.0 checkpoint, inference repository, or license was present on August 8.
“Not published” is a dated status, not a prediction. Before planning local deployment, verify all of the following on a first-party page:
- downloadable model weights and checksums;
- inference code and supported runtime versions;
- VRAM and hardware guidance;
- a license covering the intended commercial or research use;
- model-card notes on training, safety, and known limitations.
A hosted API, a site with “Wan 3.0” in its domain, or a downloadable file from an unknown uploader does not establish an official open-weight release.
Claims That Remain Unverified
The official API now establishes a 1080P ceiling and regional list prices, but it does not establish 4K output, a parameter count, one global price, or identical access across every account. Those claims remain outside the verified Wan 3.0 specification table. Preview and invitation-only labels also mean that a documented endpoint should not be described as unrestricted general availability.
The same rule applies to comparisons. An edited showcase cannot prove that Wan 3.0 beats Seedance, Kling, Veo, or MiniMax across prompt adherence, continuity, latency, safety, and cost. A useful comparison holds the input assets and acceptance criteria constant, then reports every attempt rather than a selected winner.
A Production Test for Motion Comics
For a real adoption decision, use 12 to 20 shots drawn from the work you publish. Include dialogue, full-body movement, hand-to-object contact, two-character staging, vertical framing, a camera move, and a return to an earlier scene.
Score each output on:
| Measure | What to record |
|---|---|
| Reference fidelity | Face, costume, prop, palette, and background errors |
| Motion and physics | Contact, weight, cause and effect, and unwanted morphing |
| Prompt adherence | Required action, framing, camera path, and forbidden elements |
| Continuity | Whether a later shot still matches the approved character and scene |
| Operations | Queue time, generation time, failures, retries, and output retention |
| Economics | Billed cost and human review time per approved second |
Keep the exact prompt, references, provider, region, and model settings beside every result. Without that record, a future retest cannot tell whether the model improved or the test simply changed.
Wan 3.0 and AniKuku
AniKuku currently uses an Alibaba Cloud Wan 2.6 workflow for selected video-generation tasks. AniKuku does not currently claim a Wan 3.0 integration. Adoption would follow the production test above and would also require stable API fields, failure recovery, moderation, retention, and billing behavior.
The surrounding workflow is already provider-aware: scripts become shots, approved character and scene references stay attached to the project, and failed clips can be regenerated without rebuilding the episode. See the AniKuku production workflow and finished examples for how those pieces fit together.
FAQ
Is Wan 3.0 officially released?
Yes. Alibaba Cloud Model Studio listed wan3.0-video on August 6, 2026, and Alibaba Cloud describes it as being in public beta.
What is the Wan 3.0 release date?
The official Model Studio release-log date is August 6, 2026. Account and regional availability can still differ.
What can Wan 3.0 generate?
The official entry confirms text-to-video, image-to-video, first-frame, first-and-last-frame, and reference-to-video workflows, with video up to 30 seconds.
How much does the Wan 3.0 API cost?
The documented Beijing rates are ¥0.30/¥0.60/¥1.20 per output second for 480P/720P/1080P. Singapore rates are ¥0.37471/¥0.74942/¥1.49884. The model is still marked preview or invitation-only, so verify access and the live price in your account before budgeting.
Does Wan 3.0 support 4K?
No. The current API reference documents 480P, 720P, and 1080P, not 4K. Do not treat a reseller page or YouTube title as an official specification.
Can I download Wan 3.0 weights?
No official Wan 3.0 weights were found in the Wan-AI Hugging Face or Wan-Video GitHub organizations on August 8, 2026.
Update Log
- August 8, 2026: Added the official preview API, Beijing and Singapore pricing, resolution options, duration rule, reference limits, audio defaults, and access caveats after Alibaba Cloud published the corresponding documentation.
- August 8, 2026: Published after checking the Alibaba Cloud release log, official public-beta video, official Wan GitHub and Hugging Face organizations, and current YouTube demos.
Official Sources
- Alibaba Cloud Model Studio newly released models
- Alibaba Cloud Wan 3.0 video generation API reference
- Alibaba Cloud Model Studio pricing
- Alibaba Cloud Wan 3.0 public-beta video
- Official Wan product site
- Official Wan-AI Hugging Face organization
- Official Wan-Video GitHub repositories