Wan 3.0 Reality Check: What Actually Shipped

Search for Wan 3.0 and you will find spec sheets: native 4K output, thirty-second clips, built-in sound, a six-shot AI Director mode. You will also find, at the time of writing, no Alibaba announcement, no Hugging Face weights and no QwenCloud listing behind any of it. The official Wan open-weight line still stops at Wan2.2, and that gap between what is circulating and what is verifiable is the entire story here.

This matters beyond one video model. A confident, well-designed spec page ranks. It gets screenshotted into decks, quoted in procurement conversations, and repeated by tools that summarise the web without checking whether anything underneath it is real. By the time a planning assumption reaches a production schedule, the unsourced claim has usually lost its hedge.

So this post does two things. First, it reports exactly what we checked and what came back — the vendor’s own model org, the vendor’s own paid marketplace, the derivative-artifact trail a real release leaves behind. Second, it turns those checks into a reusable method you can point at the next rumoured model, because there will be one, and it will arrive with a better-looking landing page than this one had.

The short version: at the time of writing, Wan 3.0 has not been announced or released. We did not take that from a tracker site. We queried the Hugging Face API for every model published by the official Wan-AI organisation, fetched Alibaba’s own QwenCloud marketplace page raw, and swept the community upload firehose for anything built on a Wan 3 base. All three came back consistent, and all three came back negative.

The Wan-AI org holds 24 models. The newest is Wan-AI/Wan-Dancer-14B, created July 10, 2026. The flagship open weights stop at Wan2.2 — the T2V-A14B, I2V-A14B, TI2V-5B, Animate-14B and S2V-14B variants, all Apache 2.0 licensed. There is no Wan3, no Wan-3.0, no wan3 entry anywhere in the organisation.

Alibaba’s own marketplace agrees. The featured, currently-sold Wan video product on QwenCloud is labelled Wan-T2V, model ID wan2.7-t2v, at a marketplace list rate of $0.10 to $0.15 per second of output. No Wan 3.0 product appears anywhere on that page. If you want the fuller picture of what Alibaba is actually shipping in video right now, we covered the HappyHorse 1.1 and Wan 2.7 stack when it landed.

Most coverage of a rumoured model cites the same handful of checks and shows none of the output. That is the difference worth insisting on: a claim that something does not exist is only as good as the query behind it, and a query you cannot see is an assertion. Here is the sequence — three checks plus a fourth corroboration pass — what each step is actually good for, and what it returned in this case.

Check 04 deserves its caveat in the open, because it is the one claim in this post we did not personally re-run. The no-Wan-listing finding comes from the aggregator sweep we ran for this week’s batch of coverage rather than from a fresh query in this pass. It is corroboration, not a primary — which is exactly the distinction the rest of this post is about.

There was real activity in the Wan ecosystem in July 2026, and it is worth separating from the rumour because it is the part with evidence behind it. Three names circulate together — Wan-Dancer, WanSong and Wan-Streamer — and roundups frequently describe all three as shipped Apache 2.0 models. That framing does not survive a look at the org.

Only Wan-Dancer-14B is an actual open-weights repository. WanSong and Wan-Streamer surfaced as arXiv papers; we found no public weights or license for either under the official organisation at the time of writing. The precise count is one open-weight release plus two research papers, not three models — a small correction that happens to sharpen the point, because a research cluster and a product launch are very different signals.

The cluster is genuinely tight. Wan-Dancer’s repository was created July 10 and both the WanSong and Wan-Streamer papers landed July 16 — a six-day span. Independent analysis published August 4 notes that WanSong and Wan-Streamer share a hybrid-MMDiT backbone architecturally distinct from Wan 2.2’s video-only DiT, and reads the timing as coordinated groundwork for some future unified release. That reading is reasonable. It is also an inference: at the time of writing Alibaba has stated no such plan publicly, and a shared backbone across two papers is evidence of a research direction, not of a release date.

This is the part most rumour coverage skips, and it is the part with actual signal in it. A lab publishing three artifacts in six days on a common architecture is doing something. What it is doing cannot be read off a spec sheet that nobody sourced.

The squatter sites publish a “Wan Version Lineup” table that treats 2.5, 2.6, 2.7 and 3.0 as a tidy ladder of stable, usable releases. Below is the same row structure with every cell checked against the vendor’s own surfaces instead. The shape of the table is the argument: the ladder is real up to a point, and then it stops being a ladder.

Read the middle band again, because it is the one that trips people up. Wan 2.5, 2.6 and 2.7 are not fake — they exist as commercial products. What does not exist is the downloadable Apache 2.0 build that several pages advertise. A version number being real and a version number being open are separate questions, and conflating them is the single most common failure mode in this corner of the market.

The circulating Wan 3.0 specification is not one document. It is an accumulation. One domain, wan30ai.com, markets Wan 3.0 as the latest release with native 4K video output and publishes a version lineup implying that 2.5, 2.6, 2.7 and 3.0 are all stable and usable today. A second, wan30.net, adds harder numbers: clips of up to thirty seconds, a six-shot AI Director mode, an Identity Lock feature, twelve reference assets and lip-sync across twelve languages. A community wiki adds a specific parameter count and a mid-2026 window on top.

Every one of those numbers is presented flatly, as specification. Not one of those pages links to an Alibaba document, blog post, model card or filing for any of them. That is the tell, and it is available to any reader in about ten seconds: follow the citation. If the citation is another page in the same cluster, you are reading a mirror, not a source.

The internal contradictions are almost as loud. The same ecosystem that lists Wan 2.5, 2.6 and 2.7 as stable, downloadable releases contains sites whose own explainers concede that none of those versions ever received public weights. Both statements cannot be true. We ran into the same shape of problem with a launch that shipped confident claims and no benchmarks — a real release with unverifiable numbers is a different problem from a fabricated release, but the reader-side discipline is identical.

The most instructive part of this story is not the fake spec sheet. It is watching a single unverified date appear at successively higher levels of stated confidence, until the hedge falls off.

Rung one is honest. A Hugging Face-hosted community blog post published July 29 — written by a user account, not by an official Wan or Alibaba channel — catalogues Wan-Dancer, WanSong and Wan-Streamer as ecosystem signals rather than proof of an imminent Wan 3.0, and explicitly labels an August 6, 2026 release date as an unverified leak with no Alibaba corroboration. That is a well-behaved piece of writing. It says what it knows and marks what it doesn’t.

Rung two is where it goes wrong. In the same search sweep we found a tracker page asserting something far more specific: that QwenCloud’s model release log carries a wan3.0-video entry dated August 6, described as an invite-only test. We could not corroborate that claim anywhere. It is directly contradicted by the QwenCloud page itself, which we fetched at the time of writing and which lists wan2.7-t2v as the featured Wan video product with no Wan 3.0 entry. We are recording that claim here as an unverified tracker assertion we could not stand up — not as a lead, and certainly not as news.

Notice the difference between the rungs. The tracker page is undated, so we cannot establish which claim came first or whether either derived from the other. What is visible is the same claim one rung up, hardened: a leak flagged as unverified on one page, a dated entry in a named vendor system on the other, and no new evidence attached to either. Rung three writes itself: downstream summaries repeat the specific version, cite the tracker, and drop the qualifier entirely. That is how a planning-grade fact gets manufactured out of nothing, and it is why the only durable defence is the citation chain rather than the confidence of the prose.

Rumours do not attach themselves to arbitrary version numbers. They attach where there is unresolved expectation, and Wan has a specific history that manufactures it. Wan 2.1 and Wan 2.2 were genuinely Apache 2.0 and genuinely adopted — not gesture-open, but downloaded, quantised, fine-tuned and shipped by real people. That built a constituency with a reasonable belief about how the next release would arrive.

The gap in that chart is the point. The Wan2.2 diffusers repository has been downloaded roughly forty-one times as often as Wan-Dancer-14B and about four times as often as the Wan 2.1 flagship — partly because it is newer than 2.1 and far older than Wan-Dancer, and partly because it is the version people actually build production pipelines on. Adoption of that size is what makes a closed follow-up feel like a withdrawal rather than a neutral product decision.

Then 2.5 arrived without weights. So did 2.6. So did 2.7. A live issue on the official Wan2.2 GitHub repository, opened September 24, 2025 and titled “Thank you Alibaba for deceiving and using the open source community,” documents that frustration in the plainest possible terms. That thread predates this post by just over ten months, so read it as precedent rather than as current news — but it explains why every subsequent version number arrives pre-loaded with a question the vendor has not answered.

The pattern is not unique to video. We saw the same dynamic when Qwen3.8-Max shipped with open weights still pending: a headline capability release, an open-weights promise held in reserve, and a community filling the silence with speculation. When a lab leaves the openness question open, the market answers it — usually wrongly, and always confidently.

Looking forward, the honest projection is narrow. The July research cluster is real and shares a backbone, which is the kind of thing that sometimes precedes consolidation and sometimes does not. What can be said without guessing is that the next flagship will arrive into an audience primed to assume open weights, and that the gap between that assumption and the release terms is exactly where the next round of fabricated spec sheets will grow. The defence is not better rumour-reading. It is a habit of checking the org before the blog.

None of this is a reason to stall. The practical position is straightforward: plan against artifacts that exist, describe status accurately to whoever is depending on you, and keep the watch cheap enough that you find out from the source rather than from a summary.

For teams whose video output is paid creative rather than research, the release-timing question matters less than it appears. What moves performance is the iteration loop around the asset, not the generation model’s version number, which is the frame we bring to paid media engagements. Two confirmed releases from the past week are more useful planning inputs than any unreleased flagship: FLUX 3 Video reaching general availability with twenty-second native-audio clips and Seedance 2.5’s one-take video launch. Both shipped. Both can be tested this afternoon.