Wan 3.0 Release Date: Confirmed Aug 24 Launch — Now #2 Overall on OpenArt Arena

On 24 August 2026, Alibaba Cloud confirmed the official launch of Wan 3.0 — the third generation of the Tongyi Wanxiang video-model family and the successor to Wan 2.7 — 18 days after putting it into public beta on 6 August. Three and a half weeks on, the model has a second independent scoreboard, and it is a more complicated one than the first: on 15 September, OpenArt Arena, a new public benchmark scored by creative professionals on blind pairwise comparisons, opened with Wan 3.0 at #2 overall in video behind ByteDance's Seedance 2.5, and first on its video-editing board by a single point that OpenArt's own confidence intervals treat as a tie. Alibaba's official Wan account publicised the #2 placement on X in the days that followed, in the same post that repeats the three capabilities the launch has led with — 30-second takes, native audio, any reference. The ranking is OpenArt's measurement; the framing is the vendor's. The launch itself turned the beta's metered, application-gated API into a fully open one and brought Wan 3.0 to third-party creative tools: Envato's AI video tool began offering the model behind its ModelMatch engine the same week, Leonardo.Ai added Wan 3.0 to its video generator, and the model is now available natively in ComfyUI, where a single generation runs up to 30 seconds. The consumer video editor PowerDirector (CyberLink) has added it too, announced on Alibaba's official X accounts on 28 August. What the launch still does not include is open weights: the Apache-2.0 “1.3B + 14B” story circulating around Wan 3.0 appears in no Alibaba announcement, matches the spec of a 2025 model almost exactly, and cuts against the way Alibaba has actually shipped its video models for the past year. Here is what the August 24 launch changed, what the two independent boards now say, and which parts of the Wan 3.0 story you should still treat as unverified.

A sourcing note up front, because this story now has three layers. Everything in the capabilities and pricing sections is Alibaba-reported: it comes from the official public-beta announcement and Alibaba's English blog, cross-checked against Chinese press coverage. The official launch on 24 August is vendor-announced and was covered by multiple outlets the same day. The Envato, Leonardo.Ai, ComfyUI, Pixmax and PowerDirector integrations — and the third-party video-generation API platform that began serving the model in launch week — are announced by Alibaba on X and confirmed by the platforms themselves: Envato on its own blog, Leonardo.Ai in its announcement, ComfyUI on its official blog, Pixmax on its own page, PowerDirector on CyberLink's own account, and the API platform on its own model page. These are vendor-and-partner statements, not independent tests, and they are labeled as such. The launch-ad breakdown and the creator showcase described below belong in the same category: Alibaba's own account of its own production and its partners' work, published on the official Wan account on X. The third layer is independent measurement, and it is the only layer that does not come from a party selling the model. Artificial Analysis published the first non-Alibaba numbers on 3 September. OpenArt Arena published the second on 15 September, and other boards disagree with both — the divergence is the subject of its own section below. One caveat applies to the newer result specifically: the #2 placement reached most readers through Alibaba's own X account, and OpenArt Arena is run by a company whose model selector the arena is built to feed, which makes it an independent measurement carrying a commercial interest rather than an academic one. There is still no model card, no parameter count, and no widely available weights to run. The open-weight claim, in particular, is not vendor-confirmed; it is a community rumor that the official launch contradicts, and it is labeled as such wherever it appears.

What Wan 3.0 actually ships

Wan 3.0 is a text, image, video and audio-to-video generation model. The headline capability, repeated in every outlet that covered the beta, is single-run generation of clips up to 30 seconds — twice the 15-second ceiling of Wan 2.7, and the longest native single segment among Chinese video models. That is what reviewers keep describing as “from generating a shot to telling a complete story”: 30 seconds is enough for a continuous camera move or a one-take scene, which shorter clips cannot do.

• Maximum length — up to 30 seconds in a single generation, double Wan 2.7's 15 seconds, with an “intelligent duration” feature that suggests a length from the prompt and a video-extension tool for chaining longer narratives.

• Inputs — text, image, video and audio, plus, for the first time in the family, documents: doc, xls, ppt, pdf, md, txt, key, pages and numbers files, and web links. A single file of up to 100 MB / 50 pages can be turned straight into a video — the “document-to-video” feature behind most of the marketing and education interest.

• Multi-reference — up to 10 images, 5 videos and 5 audio references per generation, for consistency of characters, props, audio, spatial layout and style across shots.

• Editing — instruction-based modification of visuals, plot and dialogue after generation, rather than full regeneration.

• Stated weaknesses — Alibaba acknowledges audio quality and on-screen text accuracy still need work; that is from the vendor's own announcement, not a reviewer's verdict.

Worth separating: “30-second clips” and “native 4K output” have both been floated for Wan 3.0, but the official beta announcement only commits to 30-second generation at 480p, 720p and 1080p. Native 4K appears on third-party marketing pages for the model, not on Alibaba's announcement. Treat 4K as unconfirmed.

Pricing: a metered API, and the pattern it breaks

Wan 3.0 is priced per second of generated video. Alibaba's stated API rates are 0.3 yuan (about $0.05) per second at 480p, 0.6 yuan (about $0.10) per second at 720p, and 1.2 yuan (about $0.20) per second at 1080p. A full 30-second clip therefore costs about 9 yuan ($1.50) at 480p, 18 yuan ($3.00) at 720p, or 36 yuan ($6.00) at 1080p. The August 24 launch confirmed the same per-second structure and added a limited-time 30% discount on Alibaba Cloud's Bailian and Qianwen AI platforms. That promotion runs through 23 September — which, at the time of this update, leaves five days on it — and drops a 30-second 1080p clip to about 25 yuan, per Alibaba. Chinese coverage noted the 720p rate lands around 60% of the per-second price of mainstream competitors, the pricing message Alibaba is clearly aiming at small and mid-size teams. The discount is also the one line in this post with a hard expiry: after 23 September, budget the list rates above.

The pricing model is the story, because it is a deliberate break from the family's history. Wan 2.1 and 2.2 shipped full open weights under Apache 2.0. Wan 2.5 was announced as open and never released weights. Wan 2.6 was a closed API. Wan 2.7 was only partially open. Wan 3.0 continues that line rather than the Wan 2.1 one: it is a metered, closed API — the beta's application gate lifted at the August 24 launch — and the official launch says nothing about releasing weights. Anyone weighing “should I wait for open weights” is waiting on something Alibaba has not promised and that the recent pattern argues against.

This is also where the routing economics matter. When a video model is priced this granularly — per second, per resolution — the difference between providers is usually not the model but the billing. On OrcaRouter, the 200+ models we route are billed at provider list price, passed through at 0% markup, so a vendor price cut is live on our side the same day rather than after a contract cycle. Wan 3.0 is not on any general-purpose model-routing platform today — it is reachable through Alibaba's own consoles, through third-party creative tools such as Envato, Leonardo.Ai, ComfyUI, Pixmax and PowerDirector, and through the video-generation API platform that began serving it in launch week — but the thing to check, now that the API is fully open, is whether the per-second rate you are quoted is the vendor's own rate. Our model pages show the list price, which is how the comparison should be done.

Where you can actually try it

The API is fully open as of the August 24 launch; during the beta it was application-gated. Alibaba lists these access points in the announcement:

• Model Studio (Bailian) — Alibaba Cloud's AI platform, where you apply for model testing.

• Qwen Cloud — Alibaba's AI-native cloud, currently showing limits of two concurrent jobs, 30 requests per minute and 50 tasks in the async queue.

• Wanxiang official website and the Qwen Creation PC client — the consumer-facing surfaces.

• WonderClip (Wanjing Yike), IF STUDIO and Duiyou — Alibaba's content tools, listed alongside the primary platforms.

• Qwen app — a gradual (gray) rollout rather than a full release.

• Envato's AI video tool — a third-party creative surface that added Wan 3.0 the week of the launch, selected behind the scenes by Envato's ModelMatch rather than picked per generation. Available to Envato users; announced by Alibaba and confirmed by Envato's own blog.

• Leonardo.Ai — a third-party creative platform that has added Wan 3.0 to its video generator, announced by Alibaba on X and confirmed by Leonardo.Ai in its own announcement. Available to Leonardo users as another non-Alibaba route in.

• ComfyUI — the node-based workflow platform, which added Wan 3.0 natively with its v0.33.4 update and on Comfy Cloud. A single generation runs up to 30 seconds at 480p, 720p or 1080p, and ComfyUI's Omni-Reference nodes pull in up to 20 image, video and audio references, plus documents and web pages. Announced by Alibaba on X and confirmed by ComfyUI on its own blog; available to ComfyUI and Comfy Cloud users.

• Pixmax — a third-party AI creation platform that added Wan 3.0 around the same time, announced by Alibaba's official account on 25 August and confirmed on Pixmax's own page, which pitches “cinema-grade output” and “flawless visual consistency across any art style.” It is also the home of the official account's first creator showcase: Liz Zhang of Pixmax walks through the launch concept film and the creative workflow built around the model. Vendor-announced and confirmed on the platform's own page.

• PowerDirector — the consumer video editor from CyberLink, which has added Wan 3.0 so that users can turn simple prompts into finished AI video inside the editor itself. Announced on Alibaba's official X accounts on 28 August and echoed by PowerDirector's own Japanese account; the first mainstream consumer video editor to surface the model. Vendor-and-partner-stated, not an independent test.

• Third-party video-generation API platforms — the newest route in, announced by Alibaba's official X account in launch week. The first such platform to serve the model exposes it over a simple task-based API for prompt-to-video generation, priced at a single $0.20 per second — Alibaba's list rate for 1080p output — with audio generated in the same pass (on by default), runs of 2–30 seconds, and omni-modal references: images, video clips, audio, files and web links, plus first/last-frame conditioning as an alternative input. Vendor-announced and confirmed on the platform's own model page.

As of the August 24 launch the API is fully open; the invite-only gate described in early-August coverage has been lifted. The realistic answer to “can I call this today” is now yes — through an Alibaba console, through a creative tool such as Envato, Leonardo.Ai, ComfyUI, Pixmax or PowerDirector that has integrated Wan 3.0, or through the third-party video-generation API platform that began serving it in launch week.

The release timeline: Aug 6 beta, Aug 10 showcase, Aug 24 launch

The event flagged in the tracking signals — a “Wan 3.0 launch” on 10 August in Hangzhou — needs a precise label. It exists and it is official: the “Wan 3.0 Release and AI+Marketing Special Event”, 14:00 on 10 August at the Alibaba Center in Hangzhou's Weike area, organised by the Hangzhou Industrial and Information Service Center with Weike Alibaba Center. The agenda covers the new model's technical evolution and features, AI+marketing and e-commerce solution sharing, and hands-on Wan 3.0 exercises.

But the model was already in public beta by then. The 10 August event is not the moment the model became real — that happened on 6 August — and it is not the launch either, which landed on 24 August. It is the formal, enterprise-facing showcase: Alibaba positioning Wan 3.0 to marketing and e-commerce buyers, in the same two-day consumer-plus-enterprise cadence it ran for the Qwen app. The dates matter for one reason: if you have been reading “Wan 3.0 comes out August 10,” the accurate statement is that the model entered public beta on 6 August, got its enterprise showcase on 10 August, and officially launched on 24 August. The 24 August launch is the release that counts.

The launch ad was made with Wan 3.0 — here's the workflow

The most direct confirmation that Wan 3.0 has genuinely shipped is that Alibaba used it for its own launch campaign. Alibaba's official Wan account on X (@Alibaba_Wan) published a full breakdown of the Wan 3.0 launch ad — the concept, the character sheets, the prompts used, and even the generations that failed. To be precise about what this is: it is the vendor's own account of its own production, not an independent test, and it is the strongest evidence the vendor can offer that a closed, API-only model is doing real production work. The breakdown's existence is itself a shipping signal — Alibaba trusted the model with its own marketing.

Alibaba has since pushed the same evidence outward. The official account's creator series — “How do creators actually use Wan 3.0 in real projects — beyond pressing generate?” — opened with Liz Zhang of Pixmax_ai, discussing the launch concept film, building a full creative workflow around the model, and what is next for the platform. This is the same category of evidence as the launch-ad breakdown: a vendor-published conversation with a partner, not an independent test, but a concrete account of the model doing production work outside Alibaba's own marketing.

The workflow the breakdown describes is built around the family's hardest problem: keeping a character recognizable across multiple shots inside a single 30-second run. Alibaba's account centers on character sheets that anchor identity across shots. Community analysts who have reverse-engineered the official examples describe the same discipline in structural terms: bind each reference asset to a named subject in the prompt instead of describing it once; set global rules that hold for the whole clip, including an explicit “do not” list for motion the model otherwise defaults to; and write the timeline as timestamped beats rather than atmosphere. Those structural findings are community work on the official sample set, not vendor guidance, but they line up with what Alibaba shows.

The detail that sets the breakdown apart is that it shows the misses. Alibaba publishes generations that failed — the identity that drifted, the beat that fell apart — rather than only the frames that landed. That is an honest look at where the model still struggles and a practical norm for anyone building on it: first-pass failures are part of the workflow, not a sign the model is broken.

The open-weight rumor, and why it's unverified

Around Wan 3.0 you will find a specific claim: that it ships open weights under Apache 2.0 in two sizes — a 1.3B consumer checkpoint and a 14B professional one — and that the weights land on Hugging Face around 10 August. This is the claim to be most careful about, because it is the most widely repeated and the least supported.

• It is not in any Alibaba announcement. The official public-beta posts describe a paid, metered API. No license, no weight release, no Hugging Face or GitHub repository appears anywhere in Alibaba's own material.

• The numbers look recycled. The 1.3B + 14B pairing is the exact spec of Wan 2.1, released in February 2025. An independent teardown of the Wan 3.0 rumors published on 4 August — before the beta announcement — flagged the “1.3B + 14B, Apache 2.0” tables as unsourced and consistent with Wan 2.1 facts rather than anything Alibaba has said.

• It cuts against the trend. Open weights in the Wan family stop at Wan 2.2 (August 2025). Wan 2.5, 2.6 and 2.7 never had core weights published; what Alibaba did open in 2026 are peripheral models — a music-to-dance model, a technical report — while the core video line stayed closed. An open Wan 3.0 would reverse that pattern, which is why the skeptical read treats it as “possible, unpromised, and against the recent trend.”

• The 10 August weights rumor is unverified. It originates from a manual tracking note, not from Alibaba or any traceable source, and the 10 August event agenda is about marketing and e-commerce, not about a release on Hugging Face.

None of this proves open weights will not happen — a vendor can change course, and the second arena result arriving this week changes nothing about it either way. But every element of the claim currently fails a primary-source check, and the official launch describes the opposite distribution model. Until Alibaba publishes a repository or a license, the accurate label for Wan 3.0 open weights is: unconfirmed, and contradicted by the official announcement.

Two independent boards now, and they do not agree

Until the start of September, every quality claim about Wan 3.0 came from Alibaba. That changed on 3 September, when Artificial Analysis — which ranks video models from blind head-to-head votes on real generations rather than vendor-supplied numbers — added the model to its boards. Wan 3.0 debuted at #1 on the Video Editing Leaderboard (With Audio): Elo roughly 1,189 from a few thousand votes, a clear margin over MiniMax H3 (about 1,129) and Gemini Omni Flash (about 1,124). On Text to Video with Audio it landed a close #2 at about 1,237, effectively tied with Gemini Omni Flash at the top, and AA also lists it #5 on Image to Video with Audio. The boards show the generational jump in the family too: Alibaba's previous model, Wan 2.7, sits #5 on the same video-editing board at about 1,077.

The second board arrived on 15 September. OpenArt Arena is a new public benchmark from the creative-tool company OpenArt, and it is built differently from the video arenas that came before it: rather than one overall ranking, it scores models per production job — overall, ads, film, animation, motion design, video editing, lip sync — from blind side-by-side comparisons judged by a panel of creative professionals and roughly a thousand self-selected “tastemakers,” aggregated with a Bradley-Terry model and reported with 95% confidence intervals. Its opening video board puts Wan 3.0 at #2 overall with 1,004 points, behind Seedance 2.5 at 1,081 and just ahead of Seedance 2.0 at 1,000. On the video-editing board Wan 3.0 takes first place — by one point, 1,034 to 1,033, a margin OpenArt itself treats as inside the noise. Alibaba's official Wan account announced the #2 placement on X in the days after the launch, in a post that repeats the three capabilities the launch has led with: 30-second takes, native audio, any reference.

So the two boards agree that Wan 3.0 belongs at the front of the field, and disagree about where. Artificial Analysis has it first in video editing and a whisker off first in text-to-video with audio; OpenArt Arena has it second overall and joint-first on editing. Boards from the same weeks are less generous still — Arena, the platform formerly known as LMArena, placed Wan 3.0 third in its Week 36 video standings, and BenchmarkList's video-generation arena also had it third, both behind Gemini Omni Flash variants. Three caveats matter before anyone treats a single number as settled. The arenas disagree because they ask different questions of different voter pools: OpenArt's judges are creative professionals voting on production tasks, Artificial Analysis' are a broader crowd voting on raw output, and neither is a controlled lab benchmark. OpenArt's own snapshots have already drifted — the top three on its overall video board have been reported at 1,081 / 1,004 / 1,000 in one capture and 1,125 / 1,047 / 1,044 in a later one, the same ordering with different gaps. And OpenArt is not a disinterested party: the Arena is designed to feed its own model selector, so it has a commercial interest in the models it ranks. That does not make the numbers wrong; it makes them a vendor-adjacent measurement rather than an academic one, and it is the same standard this post applies to Alibaba's own figures.

The practical consequence is narrower than the headline. Nothing in the OpenArt result changes what Wan 3.0 costs or what it can do — the per-second rates, the 30-second ceiling, the document input and the multimodal references belong to the same model Alibaba launched on 24 August. What the two boards change is how much confidence the quality claim deserves, in both directions. The model is no longer a vendor-claimed unknown: two independent panels put it in the top two or three of the field. It is also no longer the unchallenged leader the earlier reading implied: on at least two other boards a Gemini Omni model sits ahead of it, and its video-editing crown rests on a one-point margin on one board and a roughly 60-point margin on another. Price is a large part of why the #2 placement still reads as a win. Third-party coverage of the arena puts Wan 3.0 at roughly $0.50 per five-second clip against about $1.16 for Seedance 2.5 — figures that line up with Alibaba's 720p list rate (about $0.10 per second) rather than its 1080p one, so check the resolution before repeating them. At 1080p the same five seconds costs about $1.00.

The release-date question, answered

For the searches this post targets, the summary is short:

• When did Wan 3.0 release? Public beta: 6 August 2026, confirmed by Alibaba. Enterprise showcase: 10 August 2026 in Hangzhou. Official launch: 24 August 2026. First independent scoreboard: 3 September 2026. Second independent scoreboard: 15 September 2026.

• Is it open source? No confirmed weights. The Apache 2.0 / 1.3B + 14B claim is unverified, and the official launch is API-only.

• What does it cost? 0.3 / 0.6 / 1.2 yuan per second at 480p / 720p / 1080p; about 36 yuan ($6.00) for a 30-second 1080p clip, or about 25 yuan while the 30% launch promotion runs through 23 September.

• Can I use it? Yes. The API is fully open as of 24 August — through Model Studio or Qwen Cloud, the consumer surfaces listed above, third-party creative tools such as Envato's AI video tool, Leonardo.Ai, ComfyUI, Pixmax or PowerDirector, or a third-party video-generation API platform that began serving it in launch week.

• Where does it rank? #2 overall in video on OpenArt Arena (1,004, behind Seedance 2.5 at 1,081); #1 in video editing on both OpenArt Arena and Artificial Analysis, by one point on the former and about 60 points on the latter; #3 on Arena and BenchmarkList in the same period.

Questions people are actually asking

Is Wan 3.0 actually released, or is it still a rumor?

It is fully released: the public beta went live on 6 August, the enterprise showcase ran on 10 August, and the official launch landed on 24 August. Two independent leaderboards have since scored it. Any coverage that still calls Wan 3.0 a rumor, or an unreleased beta, is stale — including “not yet announced” analyses that were accurate before 6 August.

Will Wan 3.0 get open weights?

Nothing in Alibaba's announcement says so, the widely circulated 1.3B + 14B spec matches Wan 2.1 rather than anything new, and the family's trajectory since Wan 2.2 has been toward closed, API-only video models. Treat “open Wan 3.0” as a rumor until a repository or license appears; the recent pattern is the strongest evidence against it.

How much does a 30-second Wan 3.0 video cost?

At the stated API rates, about 36 yuan ($6.00) at 1080p, 18 yuan ($3.00) at 720p and 9 yuan ($1.50) at 480p, per clip. These are Alibaba's list prices, per second, and the API is now fully open. A 30% launch promotion runs through 23 September on Bailian and Qianwen AI, which takes the 1080p figure to about 25 yuan.

Is the 30-second generation real, and what's the catch?

The 30-second single-run generation is the feature Alibaba is leading with, and it is the longest native single segment among Chinese video models, doubling Wan 2.7. The caveats are the vendor's own: audio quality and on-screen text accuracy are acknowledged weak points. On quality, there are now two independent arenas rather than one, and they point in slightly different directions — Artificial Analysis put Wan 3.0 #1 in video editing (with audio) and a close #2 in text-to-video with audio, while OpenArt Arena puts it #2 overall and joint-first on video editing by a single point. Both measure overall output quality from blind votes rather than testing the 30-second ceiling specifically, and neither is a controlled lab benchmark.

Has anything real actually been made with Wan 3.0?

Yes — the launch ad itself. Alibaba's official Wan account on X says the Wan 3.0 launch campaign was produced with the model and has published a full breakdown of the work: the concept, the character sheets, the prompts, and the failed generations. It is a vendor-stated example rather than an independent test, but it is genuine production usage of the closed API by the people who sell it. Alibaba's official account has also showcased outside creators doing real work with the model: an interview with Liz Zhang of Pixmax_ai about the launch concept film and the creative workflow around Wan 3.0. Same label — vendor-stated, selected by Alibaba — but it moves the production-use evidence beyond Alibaba's own marketing.

The bottom line

Wan 3.0 is real, it is launched, and it is priced aggressively — 30-second generation, document-to-video, per-second rates that undercut the mainstream competition at 720p, and third-party creative routes in through Envato's AI video tool, Leonardo.Ai, ComfyUI, Pixmax and PowerDirector, plus a video-generation API platform serving the model at a flat $0.20 per second. The most direct proof that it is shipping as production software is still the launch campaign itself: Alibaba says the Wan 3.0 launch ad was made with the model and has published the full breakdown, character sheets and failed generations included. What has changed since the launch is that the quality question no longer belongs to Alibaba alone. Two independent arenas have now scored the model — Artificial Analysis on 3 September and OpenArt Arena on 15 September — and the honest reading of the two together is a top-three video model with a genuine claim on first place in editing, rather than the undisputed leader a single board suggested. The open-weight claim, meanwhile, remains a recycled-spec rumor that the official launch contradicts.

Two things are worth watching, and both are datable. The 30% launch promotion on Bailian and Qianwen AI ends on 23 September, which is the point at which the pricing argument has to stand on list rates alone. And the arenas are young: OpenArt Arena is days old and already drifting between snapshots, so a second capture of the same board in a few weeks will say more than the launch numbers do. What would settle the rest is unchanged — a model card or parameter count would answer the architecture questions, a repository or license would make the open-weight rumor real, and a controlled lab benchmark, as opposed to arena Elo, would turn the quality claims into fully checkable facts. As of this update, the accurate one-line summary of Wan 3.0 is: a confirmed, aggressively priced launch from 24 August that two independent boards now place at the front of the video field — #2 overall and joint-first in editing on OpenArt Arena, #1 in video editing on Artificial Analysis — reachable through third-party creative tools and a third-party video-generation API platform, with open weights that the evidence says are unlikely to follow.