As of 29 July 2026, Wan 3.0 does not exist. Alibaba publishes no model, repository or API id under that name. The newest Wan you can call today is wan2.7, API-only. The newest Wan with open weights is Wan 2.2.
Now the interesting part: why you searched for it.
The promise attached to the words "Wan 3.0" is the most valuable thing anyone in video could ship right now — thirty seconds, generated in a single pass, at 4K. Not eight seconds. Not six five-second clips stitched together with a face that drifts at every seam. One continuous take long enough to hold a scene: an ad that plays start to finish, a product demo with a beginning and an end, a story beat that lands without a cut.
That is a real step change in format economics, and it is worth wanting. It is also, today, a claim rather than a product. Below are the receipts.
1.Is Wan 3.0 real, or is it a rumour?
Wan 3.0 is not real as of 29 July 2026. There is no Wan 3.0 repository, no Wan 3.0 weights, and no Wan 3.0 API id on Alibaba Cloud. Every page describing it is describing something Alibaba has not published.
Three checks, all first-party, all repeatable in about ninety seconds:
The code. The Wan-Video GitHub organization shows "Showing 5 of 5 repositories": Wan2.2, Wan2.1, Wan-Dancer, Wan-skills, and a fork of diffusers. The highest-numbered repo in the org is Wan2.2. There is no Wan2.5, no Wan2.6, no Wan2.7 — and no Wan3.
The weights. The Wan-AI HuggingFace organization lists 24 models. The newest numbered line published there is Wan2.2 — Wan2.2-Animate-14B, Wan2.2-S2V-14B, Wan2.2-T2V-A14B, Wan2.2-TI2V-5B. Searching HuggingFace for "wan3" returns only unrelated third-party uploads (Mohammedf3/Wan3, Danielbsittler/wan3, Himmdy/Wan3D, intchous/wan3). None are from Wan-AI.
The API. On Alibaba Cloud Model Studio's model list (page last updated 15 July 2026), no model id begins with wan3. The newest Wan id listed is wan2.7-image-pro.
Absence of a repo is not proof a model will never exist; Alibaba could announce Wan 3.0 tomorrow. But today, anyone publishing a parameter count, duration ceiling or price for Wan 3.0 has no first-party source behind it.
2.What is the newest Wan model you can actually use today?
The newest usable Wan is wan2.7, served through Alibaba Cloud Model Studio and third-party hosts. Its documented ceiling is 15 seconds at up to 1080P. The newest Wan you can download and run yourself is Wan 2.2, released under Apache 2.0.
That split — API-only versus downloadable — is where the confusion starts. Versions 2.5, 2.6 and 2.7 exist as API model ids only; they do not appear in the GitHub org or the HuggingFace org. Version 2.2 and below exist as both.
The whole lineup, as documented on 29 July 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, Replicate |
| Wan 2.2 | Yes — 4 models on HuggingFace | "Licensed under the Apache 2.0 License." | 5s at 480P and 720P (T2V-A14B) | Self-hosted, Model Studio, Replicate, WaveSpeed |
| Wan 2.5 | Not in GitHub or HF org | n/a | wan2.5-t2v-preview: "Valid values are 5 and 10." | Model Studio, Replicate, WaveSpeed |
| Wan 2.6 | Not in GitHub or HF org | n/a | 2 to 15 seconds, default 5 | Model Studio, WaveSpeed (absent from Replicate) |
| Wan 2.7 | Not in GitHub or HF org | n/a | t2v and i2v: 2–15s. videoedit: 2–10s | Model Studio, Together AI, Replicate, WaveSpeed |
| Wan-Dancer-14B | Yes — HuggingFace | apache-2.0 | "Minute-scale" music-to-dance (card claim) | Self-hosted |
| Wan 3.0 | No repo, no weights, no API id | — | — | Nowhere |
Two details worth pulling out. First, wan2.7-videoedit accepts 2 to 10 seconds — a narrower range than text-to-video, per the video editing API reference. Editing is harder than generating, and Alibaba's own limits say so.
Second, Wan 2.7 image-to-video supports three modes: first-frame-to-video, first-and-last-frame-to-video, and video continuation. Continuation is how you get past 15 seconds. It is not native single-pass duration, and that difference matters enormously.
You don't need Wan 3.0 to ship this week
The thing people are actually waiting for is a 30-second cut that holds — same face, same wardrobe, same light, all the way through. That does not require an unreleased model. It requires a shot plan and an engine that keeps your look consistent across takes.
Imagera's cinematic video studio runs that workflow now, priced in credits instead of per second of output, so an experiment that misses costs you a retry rather than a bill. New engines are added as they become available — and the accounts already producing are the ones ready the day a longer ceiling lands.
3.Where can I run Wan 2.7 outside Alibaba Cloud?
Wan 2.7 is served by at least three third-party hosts. Together AI runs it as Wan-AI/wan2.7-t2v, with "video outputs ranging from 2 to 15 seconds" and "720P and 1080P generation".
Replicate's wan-video collection carries Wan 2.7 entries spanning text-to-video, image-to-video, reference-to-video, video editing and image generation, alongside 2.5, 2.2 and 2.1. Notably, no Wan 2.6 model appears in Replicate's collection at all, while WaveSpeed AI lists collections for 2.1, 2.2, 2.5, 2.6 and 2.7.
That inconsistency is a useful signal in itself: providers pick up Wan versions unevenly. If a page tells you a version is "available on all major providers", the provider pages are one click away.
4.What do the "Wan 3.0" pages actually claim?
They claim mutually exclusive things. Pages ranking for "Wan 3.0" variously describe a 60B dense model, a 27B mixture-of-experts model with 14B active, and a 14B dense model. Some claim 60-second output, some 30, some 10. Some say 4K, some say 1080p. They cannot all be right.
This is the clearest tell available to a reader with no inside knowledge. You do not have to evaluate anyone's credibility — only notice that the specs disagree with each other, and that none appear on a page Alibaba controls.
| Circulating claim about Wan 3.0 | What the first-party record shows (29 July 2026) |
|---|---|
| "60B parameters, dense" | No parameter count for any Wan 3.0 model exists on GitHub, HuggingFace or Model Studio. There is no Wan 3.0 model card. |
| "27B MoE, 14B active per step" | This is the published spec of Wan 2.2, verbatim, from its own README. |
| "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" | The longest documented duration for any Wan id is 15 seconds (wan2.7-t2v, wan2.7-i2v). |
| "30 seconds, single pass, 4K" | Wan 2.7's documented tiers are 720P and 1080P, including 1920*1080. No 4K tier is documented for any Wan id. |
| "Open weights on release" | Wan 2.5, 2.6 and 2.7 all shipped as API ids with no repo in the GitHub org and no model in the HF org. |
| A specific release date | Alibaba has published no announcement, changelog entry or dated post for Wan 3.0. |
To be explicit: that table compares claims against documentation. It is not an accusation against any publisher. Speculating about unreleased models is legitimate when it is labelled as speculation. The issue is that these specs circulate unlabelled, in contradictory forms, on a verification query — exactly where a reader most needs the label.
4.1Why 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 reappear attributed to an unreleased Wan 3.0, the most economical explanation is that a published spec got re-labelled upstream and copied forward. Stated neutrally: the figure is real, and it 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 most recent dated entry in that repo's Latest News is 13 November 2025 (Wan2.2-Animate-14B integrated into Diffusers), preceded by 19 September and 26 August 2025. The open-weight flagship line has been quiet; the org has not. Wan-Dancer-14B landed on 13 July 2026 under apache-2.0, "a hierarchical framework for minute-scale coherent music-to-dance generation".
5.Can any model generate 30 seconds in a single pass right now?
No model with public documentation generates 30 seconds natively. The verified ceilings are 15 seconds (Wan 2.7), 20 seconds (Sora 2), 10 seconds (Hailuo 2.3 at 768P) and 8 seconds (Veo 3.1). Longer outputs exist, but every one is produced by extension, not by a single pass.
| Model | Documented native maximum | Extension mechanics | Resolution notes |
|---|---|---|---|
| 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: Google's Veo docs, OpenAI's video generation guide, and the vendor cards above. Kling, Runway and Seedance are deliberately absent: their first-party docs did not return retrievable spec pages during this check, and second-hand blog figures are not evidence.
Read the Veo row again, because it carries the whole lesson. Veo 3.1 will hand you 148 seconds — in 21 chunks, and the moment you use extension the resolution is capped at 720p. Sora will hand you 120 seconds, in seven chunks. Longer output is real. Longer single-pass output is not what those numbers describe.
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.
6.Why is long single-pass video so hard to build?
Four documented reasons, all structural rather than a matter of effort: 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, single pass, 4K" is such an attractive claim and such a hard build. When it lands, it will land with a paper.
7.How do you ship a 30-second scene this week, without Wan 3.0?
You sequence it. A 30-second cut is not one 30-second generation — it is six shots of five seconds, planned as a scene, generated against locked references, and cut together. That workflow exists now, and it is how every polished 30-second AI spot you have watched was actually made.
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 means six chances to reroll one shot instead of rerolling thirty seconds. Directors have cut scenes this way for a century for that exact reason.
On Imagera:
- Build shots 1, 2 and 4 in the cinematic video studio — that is where prompt control and reference locking live.
- 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.
- Comparing engines before committing a client budget? Start at the video model comparison surface.
Start with shot 3, the detail insert. Smallest possible commitment, one generation, and it tells you within two minutes whether your look is right.
8.What happens the day Wan 3.0 actually ships?
New video engines are added to Imagera as they become available, and early access goes out through the studio itself. No date has been promised by anyone, including Alibaba — which is exactly why the useful move is to have your shot workflow already working before it lands, not after.
What is knowable: the last three Wan releases (2.5, 2.6, 2.7) all shipped as API endpoints first, and hosts picked them up within days. The integration path is short. Being already inside the studio, with a look you have dialled in and shot plans you have tested, is the difference between shipping on day one and starting on day one.
The cost of waiting runs one direction only. Six weeks producing 30-second cuts from sequenced shots gets you six weeks of published work and a house style. Six weeks refreshing a rumour page gets you a refreshed rumour page.
Imagera is priced in credits, not per second of output: 500 credits on Pro at $19.99, 1,500 on Business, 6,500 on Ultra — all currently shown at roughly half off. 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. Detail on the pricing page.
9.The status line, as of 29 July 2026
- Wan 3.0: no repo, no weights, no API id, no announcement. Does not exist.
- Newest Wan you can call:
wan2.7— 2 to 15 seconds, 720P and 1080P, API-only. - Newest Wan you can download: Wan 2.2, Apache 2.0, 5-second generations at 480P and 720P.
- Newest Wan release of any kind: Wan-Dancer-14B, 13 July 2026,
apache-2.0. - Longest documented single-pass output anywhere: 20 seconds (Sora 2). Everything longer is extension.
If that changes tomorrow, it changes on github.com/Wan-Video, huggingface.co/Wan-AI and Model Studio's model list before anywhere else. Those three pages take ninety seconds to check, and they outrank any blog — including this one.
Until then: the scene you were going to make with Wan 3.0 is six shots long, and you can start it today. Open the video studio.
