Updated 22 September 2026: Wan 3.0 is real. Alibaba opened it as a public beta on 6 August 2026 and formally released it on 24 August 2026. It is API-only, as wan3.0-video and the faster wan3.0-video-prime, on Alibaba Cloud Model Studio and Qwen Cloud. It renders 2 to 30 seconds in one pass at 480P, 720P or 1080P, with sound on by default. Its weights are closed. The original version of this page said on 29 July 2026 that there was no Wan 3.0. That was accurate that day and is wrong now, so we have rewritten the parts it changed.
Now the part that matters for your budget: what "thirty seconds in one pass" actually buys you.
Before August, a 30-second AI clip meant six five-second clips stitched together, with a face that drifted at every seam. Wan 3.0 removes the seams: one continuous take long enough to hold a scene. That could be an ad that plays start to finish, a product demo with a beginning and an end, or a story beat that lands without a cut. But every attempt that misses now costs you thirty seconds of render instead of five. Below are the receipts, and the workflow that keeps that cost under control.
Is Wan 3.0 real, or is it a rumour?
Yes, Wan 3.0 is real. It entered public beta on 6 August 2026 and was formally released on 24 August 2026. It has an API id, API documentation and a launch post from Alibaba Cloud. What it does not have is downloadable weights.
Three checks, all first-party or dated reporting, re-run on 22 September 2026:
The API. Alibaba Cloud's Wan3.0 video generation API reference (last updated 22 September 2026) lists two model ids: wan3.0-video, the standard version, and wan3.0-video-prime, the high-speed version. Both take the same request. Prime returns the same kind of output sooner.
The announcement. TNGlobal reported the public beta on 10 August 2026. Testers applied through Model Studio and Qwen Cloud, and the 30-second ceiling is "double the 15-second maximum of Wan2.7-Video". The Next Web covered the formal release on 24 August 2026 and noted that "the weights are closed". Alibaba Cloud's own launch post on its community blog is dated 13 August 2026.
The code. The Wan-Video GitHub organization still tops out at Wan2.2, alongside Wan-Animate-2, Wan-Dancer, Wan-skills, Wan2.1 and a diffusers fork. The org has no Wan3 repository. Alibaba's public weights organisation has no Wan 3.0 model either. Every Wan 3.0 generation runs on Alibaba-hosted infrastructure or through an API host that resells it.
So the July question has flipped. It used to be "does it exist?" Now it is "what did it ship with, and what does a 30-second attempt cost you when it misses?"
What is the newest Wan model you can actually use today?
The newest Wan is Wan 3.0, served as an API on Model Studio and Qwen Cloud. Its documented range is 2 to 30 seconds at up to 1080P. The newest general-purpose Wan you can download and run yourself is still Wan 2.2, released under Apache 2.0. The newest open-weight release of any kind is Wan2.2-Animate-2-14B, a character-animation model published on 7 August 2026 under Apache 2.0.
The API-only versus downloadable split has held since version 2.5. Versions 2.5, 2.6, 2.7 and now 3.0 exist as API model ids only. Version 2.2 and below exist as both.
The whole lineup, as documented on 22 September 2026:
| Version | Open weights? | Licence | Documented duration | Where to run it |
|---|---|---|---|---|
| Wan 2.1 | Yes — Wan-Video/Wan2.1 | Not quoted on pages checked | -plus and -turbo: "Fixed at 5 seconds and cannot be changed." | Self-hosted, Model Studio |
| Wan 2.2 | Yes — Wan-Video/Wan2.2 | "Licensed under the Apache 2.0 License." | 5s at 480P and 720P (T2V-A14B) | Self-hosted, Model Studio |
| Wan 2.5 | Not in the GitHub org | n/a | wan2.5-t2v-preview: "Valid values are 5 and 10." | Model Studio, third-party API hosts |
| Wan 2.6 | Not in the GitHub org | n/a | 2 to 15 seconds, default 5 | Model Studio, third-party API hosts |
| Wan 2.7 | Not in the GitHub org | n/a | t2v and i2v: 2–15s. videoedit: 2–10s | Model Studio, third-party API hosts |
| Wan-Dancer-14B | Yes — Wan-Video/Wan-Dancer | apache-2.0 | "Minute-scale" music-to-dance (card claim) | Self-hosted |
| Wan2.2-Animate-2-14B | Yes — released 7 August 2026 | Apache 2.0 | Character animation driven by a reference video | Self-hosted |
| Wan 3.0 | No — "the weights are closed" | n/a | 2–30s per generation. With a video input, input plus output must not exceed 30s | Model Studio, Qwen Cloud, third-party API hosts |
Two details worth pulling out. First, Wan 3.0 is one model with five jobs: text-to-video, image-to-video (first frame, or first and last frame), reference-based generation, video editing and video extension. Wan 2.7 needed a separate wan2.7-videoedit id for editing, and that id only accepts 2 to 10 seconds, per the video editing API reference.
Second, the reference inputs have hard limits. Per request, Wan 3.0 takes up to 10 images, up to 5 video clips totalling no more than 15 seconds, and up to 5 audio clips totalling no more than 15 seconds. Alibaba's launch post adds documents: one doc, xls, ppt, pdf, txt, key, pages, numbers or md file, or one web page link, capped at 100MB and 50 pages. A pitch deck can now be a reference input.
A 30-second take is only as good as the shot plan behind it
Wan 3.0 answers the length problem. It does not answer the retry problem. A take that misses the brief at second 22 costs the full thirty seconds, and you only find out after the render finishes. The fix is the same as before: prove the look on short shots, then commit the long take once the look is right.
In Imagera's video model studio, every take is priced in credits, and the exact cost of the length and resolution you pick sits on the generate button before you commit. So you choose the length you want to pay for, instead of finding out afterwards.
Where can I run Wan 3.0 and Wan 2.7?
Wan 3.0 runs on Alibaba Cloud Model Studio and on Qwen Cloud. During the beta, access was by application. The Next Web reported on 24 August that "full API availability" was rolling out, and that a consumer platform at wan.video was "promised as a members-only platform". Model Studio's Wan3.0 usage guide adds a practical catch: supported models vary by region, and the model, endpoint URL and API key must all belong to the same region, or the call fails.
Wan 2.7 is on Model Studio and several third-party inference hosts, which picked up 2.5, 2.6 and 2.7 unevenly. Some hosts skipped 2.6 entirely. Check before you build on one: if a page tells you a version is "available on all major providers", the provider pages are one click away.
Which of the circulating "Wan 3.0" claims turned out to be true?
Before launch, pages ranking for "Wan 3.0" contradicted each other. They described a 60B dense model, a 27B mixture-of-experts model with 14B active, and a 14B dense model. They claimed 60-second output, 30-second output and 10-second output. Some said 4K, some said 1080p. Now that the model has shipped, each claim can be checked against what Alibaba documents.
| Circulating claim about Wan 3.0 | What shipped (documentation checked 22 September 2026) |
|---|---|
| "60B parameters, dense" | Unverified. The API reference and launch post we read give no parameter count, and there is no public model card. |
| "27B MoE, 14B active per step" | Still the published spec of Wan 2.2, verbatim, from its own README. Nothing ties it to 3.0. |
| "14B dense" | Wan 2.2's expert models are "about 14B parameters" each. That figure belongs to 2.2's architecture. |
| "60 seconds in one pass" | False. The documented ceiling is 30 seconds per generation. Longer sequences come from the video extension task. |
| "30 seconds, single pass" | True: 2 to 30 seconds per generation. |
| "4K output" | False. The documented tiers are 480P, 720P and 1080P. |
| "Open weights on release" | False. The weights are closed, and the GitHub org has no Wan3 repository. |
| A specific release date | Public beta 6 August 2026, formal release 24 August 2026. |
To be explicit: that table compares claims against documentation. It is not an accusation against any publisher. The pre-launch pages got the headline right and most of the details wrong, which is the usual pattern with rumoured specs. Trust the API reference over any page that quotes a parameter count Alibaba has not published, including this one.
Why does "27B MoE, 14B active" sound familiar?
Because it is Wan 2.2's real, published architecture. The Wan2.2 README states it plainly: "Each expert model has about 14B parameters, resulting in a total of 27B parameters but only 14B active parameters per step."
That is a shipped 2025 model with downloadable weights, described in Alibaba's own repository. When the same numbers were attributed to Wan 3.0 before launch, the most likely explanation was that a published spec had been relabelled and copied forward. Wan 3.0 shipped without a published parameter count, so the figure still belongs to 2.2.
Two more Wan 2.2 facts from that README that rarely survive into "Wan 3.0" coverage:
- The high-compression Wan2.2-VAE "achieves a T×H×W compression ratio of 4×16×16, increasing the overall compression rate to 64 while maintaining high-quality video reconstruction." Note the axis order — 4×16×16 in time-height-width, not 16×16×4.
- TI2V-5B "supports both text-to-video and image-to-video generation at 720P resolution with 24fps and can also run on consumer-grade graphics cards like 4090", and "can generate a 5-second 720P video in under 9 minutes on a single consumer-grade GPU."
The open-weight line has not stopped. Wan-Dancer landed in July 2026 under apache-2.0, "a hierarchical framework for minute-scale coherent music-to-dance generation". Wan-Animate-2 followed on 7 August 2026, one day after the Wan 3.0 beta opened. The pattern so far: specialist models ship with open weights, and the flagship stays behind the API.
Can any model generate 30 seconds in a single pass right now?
Yes, since August 2026. Wan 3.0 documents 2 to 30 seconds per generation at up to 1080P. ByteDance's Seedance 2.5 documents 4 to 30 seconds at up to 1080p. In July, the longest native single pass we could verify was 20 seconds (Sora 2). Everything longer than that was, and outside these two models still is, produced by extension.
| Model | Documented native maximum | Extension mechanics | Resolution notes |
|---|---|---|---|
| Wan 3.0 (API) | 2–30s per generation | Video extension task; with a video input, input plus output stays within 30s | 480P, 720P and 1080P; 30fps by default |
| Seedance 2.5 (API) | 4–30s per generation | Multi-round extension | 480p, 720p and 1080p; the announced 4K did not ship |
| Wan 2.7 (API) | 2–15s; videoedit 2–10s | Video continuation mode in i2v | 720P and 1080P |
| Wan 2.2 (open weights) | 5s | None documented | 480P and 720P |
| Google Veo 3.1 | 4, 6 or 8s — must be 8 "when using extension, reference images or with 1080p and 4k resolutions" | +7s per extension, up to 20 times, "up to 148 seconds" combined | Extension is "720p" only |
| OpenAI Sora 2 / 2-pro | "16- and 20-second generations" | "up to 20 seconds" per extension, six times, "maximum total length of 120 seconds" | 1080p exports need sora-2-pro |
| MiniMax Hailuo 2.3 | 10s at 768P, 6s at 1080P; default 6s | Not documented | Values depend on model and resolution |
| Lightricks LTX-2 / 2.3 | No maximum stated in seconds | Not documented | Only rule on either card: "Frame count must be divisible by 8 + 1" |
| Tencent HunyuanVideo | --video-length 129 frames by default | Not documented | — |
Sources: Alibaba's Wan3.0 API reference, BytePlus's Seedance 2.5 API tutorial (updated 22 September 2026) for its duration and resolution limits, an AIBase report of the Volcano Engine API release dated 7 August 2026, Google's Veo docs, OpenAI's video generation guide, and the vendor cards above. We checked the Wan 3.0 and Seedance 2.5 rows on 22 September 2026. The other rows date from 29 July 2026. Kling and Runway are deliberately absent: their first-party docs did not return retrievable spec pages during the original check, and second-hand blog figures are not evidence.
Read the Veo row again, because it shows why this matters. Veo 3.1 will hand you 148 seconds, but in 21 chunks, and as soon as you use extension the resolution is capped at 720p. Sora will hand you 120 seconds in seven chunks. Wan 3.0 and Seedance 2.5 are the first two models that document a full 30-second take as one generation.
One more current-state note: Google's Gemini API video overview (updated 30 June 2026) now tells developers to "Use Gemini Omni Flash as your default model for video generation" rather than naming Veo.
Why is long single-pass video so hard to build?
There are four documented reasons, and all of them are structural: quadratic attention cost, long-context memory, error accumulation across frames, and the fidelity penalty of temporal compression. Every long-video system trades one against another.
Attention scales quadratically. Research on long-context video states it directly: "scaling diffusion transformers to generate long-context videos is fundamentally limited by the quadratic cost of self-attention" (arXiv:2508.21058). Double the clip, quadruple the attention bill.
It is a memory problem too. The same work frames it as long-context retrieval: "models must retain and retrieve salient events across a long range without collapsing or drifting." The jacket has to still be the same jacket at second 26.
Errors compound. Later work lists the failure modes as "error accumulation, attribute drift, and the limited availability of long video data" (arXiv:2606.22370). Every frame conditions the next, so small mistakes get inherited and amplified.
Compressing time costs fidelity. The obvious fix is squeezing more seconds into fewer tokens, and that has a documented ceiling: pushing "beyond 4x without increasing channel capacity poses significant challenges", and low-compression encoding of subsampled video "surpasses that of high-compression encoders applied to original videos" (arXiv:2501.05442). Wan 2.2's VAE sits at exactly 4× on the temporal axis — right at the documented edge.
That is why 30 seconds in a single pass took until August 2026 to ship, and why both models that shipped it stopped short of the 4K the rumours promised. Alibaba's own launch post says audio texture and on-screen text rendering accuracy "are improving but not yet where we want them".
How do you ship a 30-second scene this week?
Plan it as shots first, even when the final render will be one take. A single 30-second generation is now possible, but it is also the most expensive way to find out that the lighting is wrong. Block the scene as six five-second shots, lock the look on the cheap ones, then decide shot by shot whether a cut or one continuous take serves it better.
The shot plan that works, and it is deliberately boring:
- Establishing (0–5s) — wide, sets place and light. Becomes your look reference for everything after.
- Subject entry (5–10s) — person, product or car enters frame. Same key light direction as shot 1.
- Detail insert (10–15s) — macro. Texture, label, hands, badge. Cheapest to get right, highest perceived production value.
- Motion beat (15–20s) — the only shot with real camera movement. Push in or track.
- Turn (20–25s) — the reveal, the reaction, the before and after.
- Payoff (25–30s) — hold the hero frame for the end card.
Six shots give you six chances to reroll one shot instead of rerolling thirty seconds. Once shots 1, 2 and 6 hold the look, a single 30-second take of the same scene has a far better chance of landing first time, and a miss is no longer a surprise.
On Imagera:
- Build shots 1, 2 and 4 in the cinematic video studio — that is where prompt control and reference locking live.
- Ready for the continuous take? Seedance 2.5 runs in the video model studio at up to 30 seconds in one pass, with the credit cost shown before you commit.
- Working from stills? The reel makers take a photo set to a finished cut: universal reel maker, product reel maker for e-commerce, car reel maker for dealers, real estate reel maker for listings.
- Sitting on long footage already? AI video to reels cuts it into shorts — the reverse problem, often the faster win.
Start with shot 3, the detail insert. It is the smallest possible commitment: one generation, and within two minutes it tells you whether your look is right.
What changed when Wan 3.0 shipped?
Three things changed, and one did not.
Length changed. A continuous 30-second take is now a documented API call rather than a stitching job.
Inputs changed. A deck, a PDF or a web page can now steer a video, next to images, clips and audio, so a product brief can go in as the brief itself.
Openness changed. Wan 3.0 is the first Wan flagship with no downloadable weights. If your pipeline depended on self-hosting the newest Wan, it now depends on an API.
What did not change is the cost of a miss. At any per-second price, a 30-second take costs six times a five-second take, and an attempt that misses costs the same as one that lands. If you arrived with shot plans and reference sets already tested, you could use Wan 3.0 the week it opened. If you waited for the launch to start planning, you started from scratch after it.
Imagera is priced in credits, not per second of output, and every generate button shows its credit cost before you run it. Monthly plan credits reset with each cycle, while add-on credit packs never expire, so capacity you top up for a launch is still there next quarter. Details are on the pricing page.
The status line, as of 22 September 2026
- Wan 3.0: public beta 6 August 2026, formal release 24 August 2026. API-only as
wan3.0-videoandwan3.0-video-primeon Model Studio and Qwen Cloud. Closed weights. - What it renders: 2 to 30 seconds per generation, 480P, 720P or 1080P, 30fps, audio on by default.
- Newest Wan you can download: Wan 2.2 for general video, Apache 2.0, 5-second generations at 480P and 720P. Newest open release of any kind: Wan2.2-Animate-2-14B, 7 August 2026, Apache 2.0.
- Longest documented single-pass output: 30 seconds, shared by Wan 3.0 and Seedance 2.5. Everything longer is extension.
If that changes, it changes first on Model Studio's Wan3.0 API reference and github.com/Wan-Video, not on a blog. Those two pages take ninety seconds to check, and they outrank any blog — including this one.
The scene you were waiting on Wan 3.0 for is six shots long, and one of them can be a single 30-second take. Open the video studio.
